1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
4 * Written by Alex Tomas <alex@clusterfs.com>
9 * mballoc.c contains the multiblocks allocation routines
12 #include "ext4_jbd2.h"
14 #include <linux/log2.h>
15 #include <linux/module.h>
16 #include <linux/slab.h>
17 #include <linux/nospec.h>
18 #include <linux/backing-dev.h>
19 #include <linux/freezer.h>
20 #include <trace/events/ext4.h>
21 #include <kunit/static_stub.h>
25 * - test ext4_ext_search_left() and ext4_ext_search_right()
26 * - search for metadata in few groups
29 * - normalization should take into account whether file is still open
30 * - discard preallocations if no free space left (policy?)
31 * - don't normalize tails
33 * - reservation for superuser
36 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
37 * - track min/max extents in each group for better group selection
38 * - mb_mark_used() may allocate chunk right after splitting buddy
39 * - tree of groups sorted by number of free blocks
44 * The allocation request involve request for multiple number of blocks
45 * near to the goal(block) value specified.
47 * During initialization phase of the allocator we decide to use the
48 * group preallocation or inode preallocation depending on the size of
49 * the file. The size of the file could be the resulting file size we
50 * would have after allocation, or the current file size, which ever
51 * is larger. If the size is less than sbi->s_mb_stream_request we
52 * select to use the group preallocation. The default value of
53 * s_mb_stream_request is 16 blocks. This can also be tuned via
54 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
55 * terms of number of blocks.
57 * The main motivation for having small file use group preallocation is to
58 * ensure that we have small files closer together on the disk.
60 * First stage the allocator looks at the inode prealloc list,
61 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
62 * spaces for this particular inode. The inode prealloc space is
65 * pa_lstart -> the logical start block for this prealloc space
66 * pa_pstart -> the physical start block for this prealloc space
67 * pa_len -> length for this prealloc space (in clusters)
68 * pa_free -> free space available in this prealloc space (in clusters)
70 * The inode preallocation space is used looking at the _logical_ start
71 * block. If only the logical file block falls within the range of prealloc
72 * space we will consume the particular prealloc space. This makes sure that
73 * we have contiguous physical blocks representing the file blocks
75 * The important thing to be noted in case of inode prealloc space is that
76 * we don't modify the values associated to inode prealloc space except
79 * If we are not able to find blocks in the inode prealloc space and if we
80 * have the group allocation flag set then we look at the locality group
81 * prealloc space. These are per CPU prealloc list represented as
83 * ext4_sb_info.s_locality_groups[smp_processor_id()]
85 * The reason for having a per cpu locality group is to reduce the contention
86 * between CPUs. It is possible to get scheduled at this point.
88 * The locality group prealloc space is used looking at whether we have
89 * enough free space (pa_free) within the prealloc space.
91 * If we can't allocate blocks via inode prealloc or/and locality group
92 * prealloc then we look at the buddy cache. The buddy cache is represented
93 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
94 * mapped to the buddy and bitmap information regarding different
95 * groups. The buddy information is attached to buddy cache inode so that
96 * we can access them through the page cache. The information regarding
97 * each group is loaded via ext4_mb_load_buddy. The information involve
98 * block bitmap and buddy information. The information are stored in the
102 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
105 * one block each for bitmap and buddy information. So for each group we
106 * take up 2 blocks. A page can contain blocks_per_page (PAGE_SIZE /
107 * blocksize) blocks. So it can have information regarding groups_per_page
108 * which is blocks_per_page/2
110 * The buddy cache inode is not stored on disk. The inode is thrown
111 * away when the filesystem is unmounted.
113 * We look for count number of blocks in the buddy cache. If we were able
114 * to locate that many free blocks we return with additional information
115 * regarding rest of the contiguous physical block available
117 * Before allocating blocks via buddy cache we normalize the request
118 * blocks. This ensure we ask for more blocks that we needed. The extra
119 * blocks that we get after allocation is added to the respective prealloc
120 * list. In case of inode preallocation we follow a list of heuristics
121 * based on file size. This can be found in ext4_mb_normalize_request. If
122 * we are doing a group prealloc we try to normalize the request to
123 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
124 * dependent on the cluster size; for non-bigalloc file systems, it is
125 * 512 blocks. This can be tuned via
126 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
127 * terms of number of blocks. If we have mounted the file system with -O
128 * stripe=<value> option the group prealloc request is normalized to the
129 * smallest multiple of the stripe value (sbi->s_stripe) which is
130 * greater than the default mb_group_prealloc.
132 * If "mb_optimize_scan" mount option is set, we maintain in memory group info
133 * structures in two data structures:
135 * 1) Array of largest free order lists (sbi->s_mb_largest_free_orders)
137 * Locking: sbi->s_mb_largest_free_orders_locks(array of rw locks)
139 * This is an array of lists where the index in the array represents the
140 * largest free order in the buddy bitmap of the participating group infos of
141 * that list. So, there are exactly MB_NUM_ORDERS(sb) (which means total
142 * number of buddy bitmap orders possible) number of lists. Group-infos are
143 * placed in appropriate lists.
145 * 2) Average fragment size lists (sbi->s_mb_avg_fragment_size)
147 * Locking: sbi->s_mb_avg_fragment_size_locks(array of rw locks)
149 * This is an array of lists where in the i-th list there are groups with
150 * average fragment size >= 2^i and < 2^(i+1). The average fragment size
151 * is computed as ext4_group_info->bb_free / ext4_group_info->bb_fragments.
152 * Note that we don't bother with a special list for completely empty groups
153 * so we only have MB_NUM_ORDERS(sb) lists.
155 * When "mb_optimize_scan" mount option is set, mballoc consults the above data
156 * structures to decide the order in which groups are to be traversed for
157 * fulfilling an allocation request.
159 * At CR_POWER2_ALIGNED , we look for groups which have the largest_free_order
160 * >= the order of the request. We directly look at the largest free order list
161 * in the data structure (1) above where largest_free_order = order of the
162 * request. If that list is empty, we look at remaining list in the increasing
163 * order of largest_free_order. This allows us to perform CR_POWER2_ALIGNED
164 * lookup in O(1) time.
166 * At CR_GOAL_LEN_FAST, we only consider groups where
167 * average fragment size > request size. So, we lookup a group which has average
168 * fragment size just above or equal to request size using our average fragment
169 * size group lists (data structure 2) in O(1) time.
171 * At CR_BEST_AVAIL_LEN, we aim to optimize allocations which can't be satisfied
172 * in CR_GOAL_LEN_FAST. The fact that we couldn't find a group in
173 * CR_GOAL_LEN_FAST suggests that there is no BG that has avg
174 * fragment size > goal length. So before falling to the slower
175 * CR_GOAL_LEN_SLOW, in CR_BEST_AVAIL_LEN we proactively trim goal length and
176 * then use the same fragment lists as CR_GOAL_LEN_FAST to find a BG with a big
177 * enough average fragment size. This increases the chances of finding a
178 * suitable block group in O(1) time and results in faster allocation at the
179 * cost of reduced size of allocation.
181 * If "mb_optimize_scan" mount option is not set, mballoc traverses groups in
182 * linear order which requires O(N) search time for each CR_POWER2_ALIGNED and
183 * CR_GOAL_LEN_FAST phase.
185 * The regular allocator (using the buddy cache) supports a few tunables.
187 * /sys/fs/ext4/<partition>/mb_min_to_scan
188 * /sys/fs/ext4/<partition>/mb_max_to_scan
189 * /sys/fs/ext4/<partition>/mb_order2_req
190 * /sys/fs/ext4/<partition>/mb_linear_limit
192 * The regular allocator uses buddy scan only if the request len is power of
193 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
194 * value of s_mb_order2_reqs can be tuned via
195 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
196 * stripe size (sbi->s_stripe), we try to search for contiguous block in
197 * stripe size. This should result in better allocation on RAID setups. If
198 * not, we search in the specific group using bitmap for best extents. The
199 * tunable min_to_scan and max_to_scan control the behaviour here.
200 * min_to_scan indicate how long the mballoc __must__ look for a best
201 * extent and max_to_scan indicates how long the mballoc __can__ look for a
202 * best extent in the found extents. Searching for the blocks starts with
203 * the group specified as the goal value in allocation context via
204 * ac_g_ex. Each group is first checked based on the criteria whether it
205 * can be used for allocation. ext4_mb_good_group explains how the groups are
208 * When "mb_optimize_scan" is turned on, as mentioned above, the groups may not
209 * get traversed linearly. That may result in subsequent allocations being not
210 * close to each other. And so, the underlying device may get filled up in a
211 * non-linear fashion. While that may not matter on non-rotational devices, for
212 * rotational devices that may result in higher seek times. "mb_linear_limit"
213 * tells mballoc how many groups mballoc should search linearly before
214 * performing consulting above data structures for more efficient lookups. For
215 * non rotational devices, this value defaults to 0 and for rotational devices
216 * this is set to MB_DEFAULT_LINEAR_LIMIT.
218 * Both the prealloc space are getting populated as above. So for the first
219 * request we will hit the buddy cache which will result in this prealloc
220 * space getting filled. The prealloc space is then later used for the
221 * subsequent request.
225 * mballoc operates on the following data:
227 * - in-core buddy (actually includes buddy and bitmap)
228 * - preallocation descriptors (PAs)
230 * there are two types of preallocations:
232 * assiged to specific inode and can be used for this inode only.
233 * it describes part of inode's space preallocated to specific
234 * physical blocks. any block from that preallocated can be used
235 * independent. the descriptor just tracks number of blocks left
236 * unused. so, before taking some block from descriptor, one must
237 * make sure corresponded logical block isn't allocated yet. this
238 * also means that freeing any block within descriptor's range
239 * must discard all preallocated blocks.
241 * assigned to specific locality group which does not translate to
242 * permanent set of inodes: inode can join and leave group. space
243 * from this type of preallocation can be used for any inode. thus
244 * it's consumed from the beginning to the end.
246 * relation between them can be expressed as:
247 * in-core buddy = on-disk bitmap + preallocation descriptors
249 * this mean blocks mballoc considers used are:
250 * - allocated blocks (persistent)
251 * - preallocated blocks (non-persistent)
253 * consistency in mballoc world means that at any time a block is either
254 * free or used in ALL structures. notice: "any time" should not be read
255 * literally -- time is discrete and delimited by locks.
257 * to keep it simple, we don't use block numbers, instead we count number of
258 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
260 * all operations can be expressed as:
261 * - init buddy: buddy = on-disk + PAs
262 * - new PA: buddy += N; PA = N
263 * - use inode PA: on-disk += N; PA -= N
264 * - discard inode PA buddy -= on-disk - PA; PA = 0
265 * - use locality group PA on-disk += N; PA -= N
266 * - discard locality group PA buddy -= PA; PA = 0
267 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
268 * is used in real operation because we can't know actual used
269 * bits from PA, only from on-disk bitmap
271 * if we follow this strict logic, then all operations above should be atomic.
272 * given some of them can block, we'd have to use something like semaphores
273 * killing performance on high-end SMP hardware. let's try to relax it using
274 * the following knowledge:
275 * 1) if buddy is referenced, it's already initialized
276 * 2) while block is used in buddy and the buddy is referenced,
277 * nobody can re-allocate that block
278 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
279 * bit set and PA claims same block, it's OK. IOW, one can set bit in
280 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
283 * so, now we're building a concurrency table:
286 * blocks for PA are allocated in the buddy, buddy must be referenced
287 * until PA is linked to allocation group to avoid concurrent buddy init
289 * we need to make sure that either on-disk bitmap or PA has uptodate data
290 * given (3) we care that PA-=N operation doesn't interfere with init
292 * the simplest way would be to have buddy initialized by the discard
293 * - use locality group PA
294 * again PA-=N must be serialized with init
295 * - discard locality group PA
296 * the simplest way would be to have buddy initialized by the discard
299 * i_data_sem serializes them
301 * discard process must wait until PA isn't used by another process
302 * - use locality group PA
303 * some mutex should serialize them
304 * - discard locality group PA
305 * discard process must wait until PA isn't used by another process
308 * i_data_sem or another mutex should serializes them
310 * discard process must wait until PA isn't used by another process
311 * - use locality group PA
312 * nothing wrong here -- they're different PAs covering different blocks
313 * - discard locality group PA
314 * discard process must wait until PA isn't used by another process
316 * now we're ready to make few consequences:
317 * - PA is referenced and while it is no discard is possible
318 * - PA is referenced until block isn't marked in on-disk bitmap
319 * - PA changes only after on-disk bitmap
320 * - discard must not compete with init. either init is done before
321 * any discard or they're serialized somehow
322 * - buddy init as sum of on-disk bitmap and PAs is done atomically
324 * a special case when we've used PA to emptiness. no need to modify buddy
325 * in this case, but we should care about concurrent init
330 * Logic in few words:
335 * mark bits in on-disk bitmap
338 * - use preallocation:
339 * find proper PA (per-inode or group)
341 * mark bits in on-disk bitmap
347 * mark bits in on-disk bitmap
350 * - discard preallocations in group:
352 * move them onto local list
353 * load on-disk bitmap
355 * remove PA from object (inode or locality group)
356 * mark free blocks in-core
358 * - discard inode's preallocations:
365 * - bitlock on a group (group)
366 * - object (inode/locality) (object)
368 * - cr_power2_aligned lists lock (cr_power2_aligned)
369 * - cr_goal_len_fast lists lock (cr_goal_len_fast)
379 * - release consumed pa:
384 * - generate in-core bitmap:
388 * - discard all for given object (inode, locality group):
393 * - discard all for given group:
399 * - allocation path (ext4_mb_regular_allocator)
401 * cr_power2_aligned/cr_goal_len_fast
403 static struct kmem_cache *ext4_pspace_cachep;
404 static struct kmem_cache *ext4_ac_cachep;
405 static struct kmem_cache *ext4_free_data_cachep;
407 /* We create slab caches for groupinfo data structures based on the
408 * superblock block size. There will be one per mounted filesystem for
409 * each unique s_blocksize_bits */
410 #define NR_GRPINFO_CACHES 8
411 static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
413 static const char * const ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
414 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
415 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
416 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
419 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
421 static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac);
423 static bool ext4_mb_good_group(struct ext4_allocation_context *ac,
424 ext4_group_t group, enum criteria cr);
426 static int ext4_try_to_trim_range(struct super_block *sb,
427 struct ext4_buddy *e4b, ext4_grpblk_t start,
428 ext4_grpblk_t max, ext4_grpblk_t minblocks);
431 * The algorithm using this percpu seq counter goes below:
432 * 1. We sample the percpu discard_pa_seq counter before trying for block
433 * allocation in ext4_mb_new_blocks().
434 * 2. We increment this percpu discard_pa_seq counter when we either allocate
435 * or free these blocks i.e. while marking those blocks as used/free in
436 * mb_mark_used()/mb_free_blocks().
437 * 3. We also increment this percpu seq counter when we successfully identify
438 * that the bb_prealloc_list is not empty and hence proceed for discarding
439 * of those PAs inside ext4_mb_discard_group_preallocations().
441 * Now to make sure that the regular fast path of block allocation is not
442 * affected, as a small optimization we only sample the percpu seq counter
443 * on that cpu. Only when the block allocation fails and when freed blocks
444 * found were 0, that is when we sample percpu seq counter for all cpus using
445 * below function ext4_get_discard_pa_seq_sum(). This happens after making
446 * sure that all the PAs on grp->bb_prealloc_list got freed or if it's empty.
448 static DEFINE_PER_CPU(u64, discard_pa_seq);
449 static inline u64 ext4_get_discard_pa_seq_sum(void)
454 for_each_possible_cpu(__cpu)
455 __seq += per_cpu(discard_pa_seq, __cpu);
459 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
461 #if BITS_PER_LONG == 64
462 *bit += ((unsigned long) addr & 7UL) << 3;
463 addr = (void *) ((unsigned long) addr & ~7UL);
464 #elif BITS_PER_LONG == 32
465 *bit += ((unsigned long) addr & 3UL) << 3;
466 addr = (void *) ((unsigned long) addr & ~3UL);
468 #error "how many bits you are?!"
473 static inline int mb_test_bit(int bit, void *addr)
476 * ext4_test_bit on architecture like powerpc
477 * needs unsigned long aligned address
479 addr = mb_correct_addr_and_bit(&bit, addr);
480 return ext4_test_bit(bit, addr);
483 static inline void mb_set_bit(int bit, void *addr)
485 addr = mb_correct_addr_and_bit(&bit, addr);
486 ext4_set_bit(bit, addr);
489 static inline void mb_clear_bit(int bit, void *addr)
491 addr = mb_correct_addr_and_bit(&bit, addr);
492 ext4_clear_bit(bit, addr);
495 static inline int mb_test_and_clear_bit(int bit, void *addr)
497 addr = mb_correct_addr_and_bit(&bit, addr);
498 return ext4_test_and_clear_bit(bit, addr);
501 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
503 int fix = 0, ret, tmpmax;
504 addr = mb_correct_addr_and_bit(&fix, addr);
508 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
514 static inline int mb_find_next_bit(void *addr, int max, int start)
516 int fix = 0, ret, tmpmax;
517 addr = mb_correct_addr_and_bit(&fix, addr);
521 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
527 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
531 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
534 if (order > e4b->bd_blkbits + 1) {
539 /* at order 0 we see each particular block */
541 *max = 1 << (e4b->bd_blkbits + 3);
542 return e4b->bd_bitmap;
545 bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
546 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
552 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
553 int first, int count)
556 struct super_block *sb = e4b->bd_sb;
558 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
560 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
561 for (i = 0; i < count; i++) {
562 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
563 ext4_fsblk_t blocknr;
565 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
566 blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
567 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
568 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
569 ext4_grp_locked_error(sb, e4b->bd_group,
570 inode ? inode->i_ino : 0,
572 "freeing block already freed "
576 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
580 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
584 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
586 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
587 for (i = 0; i < count; i++) {
588 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
589 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
593 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
595 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
597 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
598 unsigned char *b1, *b2;
600 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
601 b2 = (unsigned char *) bitmap;
602 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
603 if (b1[i] != b2[i]) {
604 ext4_msg(e4b->bd_sb, KERN_ERR,
605 "corruption in group %u "
606 "at byte %u(%u): %x in copy != %x "
608 e4b->bd_group, i, i * 8, b1[i], b2[i]);
615 static void mb_group_bb_bitmap_alloc(struct super_block *sb,
616 struct ext4_group_info *grp, ext4_group_t group)
618 struct buffer_head *bh;
620 grp->bb_bitmap = kmalloc(sb->s_blocksize, GFP_NOFS);
624 bh = ext4_read_block_bitmap(sb, group);
625 if (IS_ERR_OR_NULL(bh)) {
626 kfree(grp->bb_bitmap);
627 grp->bb_bitmap = NULL;
631 memcpy(grp->bb_bitmap, bh->b_data, sb->s_blocksize);
635 static void mb_group_bb_bitmap_free(struct ext4_group_info *grp)
637 kfree(grp->bb_bitmap);
641 static inline void mb_free_blocks_double(struct inode *inode,
642 struct ext4_buddy *e4b, int first, int count)
646 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
647 int first, int count)
651 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
656 static inline void mb_group_bb_bitmap_alloc(struct super_block *sb,
657 struct ext4_group_info *grp, ext4_group_t group)
662 static inline void mb_group_bb_bitmap_free(struct ext4_group_info *grp)
668 #ifdef AGGRESSIVE_CHECK
670 #define MB_CHECK_ASSERT(assert) \
674 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
675 function, file, line, # assert); \
680 static void __mb_check_buddy(struct ext4_buddy *e4b, char *file,
681 const char *function, int line)
683 struct super_block *sb = e4b->bd_sb;
684 int order = e4b->bd_blkbits + 1;
691 struct ext4_group_info *grp;
694 struct list_head *cur;
698 if (e4b->bd_info->bb_check_counter++ % 10)
702 buddy = mb_find_buddy(e4b, order, &max);
703 MB_CHECK_ASSERT(buddy);
704 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
705 MB_CHECK_ASSERT(buddy2);
706 MB_CHECK_ASSERT(buddy != buddy2);
707 MB_CHECK_ASSERT(max * 2 == max2);
710 for (i = 0; i < max; i++) {
712 if (mb_test_bit(i, buddy)) {
713 /* only single bit in buddy2 may be 0 */
714 if (!mb_test_bit(i << 1, buddy2)) {
716 mb_test_bit((i<<1)+1, buddy2));
721 /* both bits in buddy2 must be 1 */
722 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
723 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
725 for (j = 0; j < (1 << order); j++) {
726 k = (i * (1 << order)) + j;
728 !mb_test_bit(k, e4b->bd_bitmap));
732 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
737 buddy = mb_find_buddy(e4b, 0, &max);
738 for (i = 0; i < max; i++) {
739 if (!mb_test_bit(i, buddy)) {
740 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
748 /* check used bits only */
749 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
750 buddy2 = mb_find_buddy(e4b, j, &max2);
752 MB_CHECK_ASSERT(k < max2);
753 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
756 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
757 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
759 grp = ext4_get_group_info(sb, e4b->bd_group);
762 list_for_each(cur, &grp->bb_prealloc_list) {
763 ext4_group_t groupnr;
764 struct ext4_prealloc_space *pa;
765 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
766 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
767 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
768 for (i = 0; i < pa->pa_len; i++)
769 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
772 #undef MB_CHECK_ASSERT
773 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
774 __FILE__, __func__, __LINE__)
776 #define mb_check_buddy(e4b)
780 * Divide blocks started from @first with length @len into
781 * smaller chunks with power of 2 blocks.
782 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
783 * then increase bb_counters[] for corresponded chunk size.
785 static void ext4_mb_mark_free_simple(struct super_block *sb,
786 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
787 struct ext4_group_info *grp)
789 struct ext4_sb_info *sbi = EXT4_SB(sb);
795 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
797 border = 2 << sb->s_blocksize_bits;
800 /* find how many blocks can be covered since this position */
801 max = ffs(first | border) - 1;
803 /* find how many blocks of power 2 we need to mark */
810 /* mark multiblock chunks only */
811 grp->bb_counters[min]++;
813 mb_clear_bit(first >> min,
814 buddy + sbi->s_mb_offsets[min]);
821 static int mb_avg_fragment_size_order(struct super_block *sb, ext4_grpblk_t len)
826 * We don't bother with a special lists groups with only 1 block free
827 * extents and for completely empty groups.
829 order = fls(len) - 2;
832 if (order == MB_NUM_ORDERS(sb))
837 /* Move group to appropriate avg_fragment_size list */
839 mb_update_avg_fragment_size(struct super_block *sb, struct ext4_group_info *grp)
841 struct ext4_sb_info *sbi = EXT4_SB(sb);
844 if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || grp->bb_fragments == 0)
847 new_order = mb_avg_fragment_size_order(sb,
848 grp->bb_free / grp->bb_fragments);
849 if (new_order == grp->bb_avg_fragment_size_order)
852 if (grp->bb_avg_fragment_size_order != -1) {
853 write_lock(&sbi->s_mb_avg_fragment_size_locks[
854 grp->bb_avg_fragment_size_order]);
855 list_del(&grp->bb_avg_fragment_size_node);
856 write_unlock(&sbi->s_mb_avg_fragment_size_locks[
857 grp->bb_avg_fragment_size_order]);
859 grp->bb_avg_fragment_size_order = new_order;
860 write_lock(&sbi->s_mb_avg_fragment_size_locks[
861 grp->bb_avg_fragment_size_order]);
862 list_add_tail(&grp->bb_avg_fragment_size_node,
863 &sbi->s_mb_avg_fragment_size[grp->bb_avg_fragment_size_order]);
864 write_unlock(&sbi->s_mb_avg_fragment_size_locks[
865 grp->bb_avg_fragment_size_order]);
869 * Choose next group by traversing largest_free_order lists. Updates *new_cr if
870 * cr level needs an update.
872 static void ext4_mb_choose_next_group_p2_aligned(struct ext4_allocation_context *ac,
873 enum criteria *new_cr, ext4_group_t *group)
875 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
876 struct ext4_group_info *iter;
879 if (ac->ac_status == AC_STATUS_FOUND)
882 if (unlikely(sbi->s_mb_stats && ac->ac_flags & EXT4_MB_CR_POWER2_ALIGNED_OPTIMIZED))
883 atomic_inc(&sbi->s_bal_p2_aligned_bad_suggestions);
885 for (i = ac->ac_2order; i < MB_NUM_ORDERS(ac->ac_sb); i++) {
886 if (list_empty(&sbi->s_mb_largest_free_orders[i]))
888 read_lock(&sbi->s_mb_largest_free_orders_locks[i]);
889 if (list_empty(&sbi->s_mb_largest_free_orders[i])) {
890 read_unlock(&sbi->s_mb_largest_free_orders_locks[i]);
893 list_for_each_entry(iter, &sbi->s_mb_largest_free_orders[i],
894 bb_largest_free_order_node) {
896 atomic64_inc(&sbi->s_bal_cX_groups_considered[CR_POWER2_ALIGNED]);
897 if (likely(ext4_mb_good_group(ac, iter->bb_group, CR_POWER2_ALIGNED))) {
898 *group = iter->bb_group;
899 ac->ac_flags |= EXT4_MB_CR_POWER2_ALIGNED_OPTIMIZED;
900 read_unlock(&sbi->s_mb_largest_free_orders_locks[i]);
904 read_unlock(&sbi->s_mb_largest_free_orders_locks[i]);
907 /* Increment cr and search again if no group is found */
908 *new_cr = CR_GOAL_LEN_FAST;
912 * Find a suitable group of given order from the average fragments list.
914 static struct ext4_group_info *
915 ext4_mb_find_good_group_avg_frag_lists(struct ext4_allocation_context *ac, int order)
917 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
918 struct list_head *frag_list = &sbi->s_mb_avg_fragment_size[order];
919 rwlock_t *frag_list_lock = &sbi->s_mb_avg_fragment_size_locks[order];
920 struct ext4_group_info *grp = NULL, *iter;
921 enum criteria cr = ac->ac_criteria;
923 if (list_empty(frag_list))
925 read_lock(frag_list_lock);
926 if (list_empty(frag_list)) {
927 read_unlock(frag_list_lock);
930 list_for_each_entry(iter, frag_list, bb_avg_fragment_size_node) {
932 atomic64_inc(&sbi->s_bal_cX_groups_considered[cr]);
933 if (likely(ext4_mb_good_group(ac, iter->bb_group, cr))) {
938 read_unlock(frag_list_lock);
943 * Choose next group by traversing average fragment size list of suitable
944 * order. Updates *new_cr if cr level needs an update.
946 static void ext4_mb_choose_next_group_goal_fast(struct ext4_allocation_context *ac,
947 enum criteria *new_cr, ext4_group_t *group)
949 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
950 struct ext4_group_info *grp = NULL;
953 if (unlikely(ac->ac_flags & EXT4_MB_CR_GOAL_LEN_FAST_OPTIMIZED)) {
955 atomic_inc(&sbi->s_bal_goal_fast_bad_suggestions);
958 for (i = mb_avg_fragment_size_order(ac->ac_sb, ac->ac_g_ex.fe_len);
959 i < MB_NUM_ORDERS(ac->ac_sb); i++) {
960 grp = ext4_mb_find_good_group_avg_frag_lists(ac, i);
962 *group = grp->bb_group;
963 ac->ac_flags |= EXT4_MB_CR_GOAL_LEN_FAST_OPTIMIZED;
969 * CR_BEST_AVAIL_LEN works based on the concept that we have
970 * a larger normalized goal len request which can be trimmed to
971 * a smaller goal len such that it can still satisfy original
972 * request len. However, allocation request for non-regular
973 * files never gets normalized.
974 * See function ext4_mb_normalize_request() (EXT4_MB_HINT_DATA).
976 if (ac->ac_flags & EXT4_MB_HINT_DATA)
977 *new_cr = CR_BEST_AVAIL_LEN;
979 *new_cr = CR_GOAL_LEN_SLOW;
983 * We couldn't find a group in CR_GOAL_LEN_FAST so try to find the highest free fragment
984 * order we have and proactively trim the goal request length to that order to
985 * find a suitable group faster.
987 * This optimizes allocation speed at the cost of slightly reduced
988 * preallocations. However, we make sure that we don't trim the request too
989 * much and fall to CR_GOAL_LEN_SLOW in that case.
991 static void ext4_mb_choose_next_group_best_avail(struct ext4_allocation_context *ac,
992 enum criteria *new_cr, ext4_group_t *group)
994 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
995 struct ext4_group_info *grp = NULL;
996 int i, order, min_order;
997 unsigned long num_stripe_clusters = 0;
999 if (unlikely(ac->ac_flags & EXT4_MB_CR_BEST_AVAIL_LEN_OPTIMIZED)) {
1000 if (sbi->s_mb_stats)
1001 atomic_inc(&sbi->s_bal_best_avail_bad_suggestions);
1005 * mb_avg_fragment_size_order() returns order in a way that makes
1006 * retrieving back the length using (1 << order) inaccurate. Hence, use
1007 * fls() instead since we need to know the actual length while modifying
1010 order = fls(ac->ac_g_ex.fe_len) - 1;
1011 min_order = order - sbi->s_mb_best_avail_max_trim_order;
1015 if (sbi->s_stripe > 0) {
1017 * We are assuming that stripe size is always a multiple of
1018 * cluster ratio otherwise __ext4_fill_super exists early.
1020 num_stripe_clusters = EXT4_NUM_B2C(sbi, sbi->s_stripe);
1021 if (1 << min_order < num_stripe_clusters)
1023 * We consider 1 order less because later we round
1024 * up the goal len to num_stripe_clusters
1026 min_order = fls(num_stripe_clusters) - 1;
1029 if (1 << min_order < ac->ac_o_ex.fe_len)
1030 min_order = fls(ac->ac_o_ex.fe_len);
1032 for (i = order; i >= min_order; i--) {
1035 * Scale down goal len to make sure we find something
1036 * in the free fragments list. Basically, reduce
1039 ac->ac_g_ex.fe_len = 1 << i;
1041 if (num_stripe_clusters > 0) {
1043 * Try to round up the adjusted goal length to
1044 * stripe size (in cluster units) multiple for
1047 ac->ac_g_ex.fe_len = roundup(ac->ac_g_ex.fe_len,
1048 num_stripe_clusters);
1051 frag_order = mb_avg_fragment_size_order(ac->ac_sb,
1052 ac->ac_g_ex.fe_len);
1054 grp = ext4_mb_find_good_group_avg_frag_lists(ac, frag_order);
1056 *group = grp->bb_group;
1057 ac->ac_flags |= EXT4_MB_CR_BEST_AVAIL_LEN_OPTIMIZED;
1062 /* Reset goal length to original goal length before falling into CR_GOAL_LEN_SLOW */
1063 ac->ac_g_ex.fe_len = ac->ac_orig_goal_len;
1064 *new_cr = CR_GOAL_LEN_SLOW;
1067 static inline int should_optimize_scan(struct ext4_allocation_context *ac)
1069 if (unlikely(!test_opt2(ac->ac_sb, MB_OPTIMIZE_SCAN)))
1071 if (ac->ac_criteria >= CR_GOAL_LEN_SLOW)
1073 if (!ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))
1079 * Return next linear group for allocation. If linear traversal should not be
1080 * performed, this function just returns the same group
1083 next_linear_group(struct ext4_allocation_context *ac, ext4_group_t group,
1084 ext4_group_t ngroups)
1086 if (!should_optimize_scan(ac))
1087 goto inc_and_return;
1089 if (ac->ac_groups_linear_remaining) {
1090 ac->ac_groups_linear_remaining--;
1091 goto inc_and_return;
1097 * Artificially restricted ngroups for non-extent
1098 * files makes group > ngroups possible on first loop.
1100 return group + 1 >= ngroups ? 0 : group + 1;
1104 * ext4_mb_choose_next_group: choose next group for allocation.
1106 * @ac Allocation Context
1107 * @new_cr This is an output parameter. If the there is no good group
1108 * available at current CR level, this field is updated to indicate
1109 * the new cr level that should be used.
1110 * @group This is an input / output parameter. As an input it indicates the
1111 * next group that the allocator intends to use for allocation. As
1112 * output, this field indicates the next group that should be used as
1113 * determined by the optimization functions.
1114 * @ngroups Total number of groups
1116 static void ext4_mb_choose_next_group(struct ext4_allocation_context *ac,
1117 enum criteria *new_cr, ext4_group_t *group, ext4_group_t ngroups)
1119 *new_cr = ac->ac_criteria;
1121 if (!should_optimize_scan(ac) || ac->ac_groups_linear_remaining) {
1122 *group = next_linear_group(ac, *group, ngroups);
1126 if (*new_cr == CR_POWER2_ALIGNED) {
1127 ext4_mb_choose_next_group_p2_aligned(ac, new_cr, group);
1128 } else if (*new_cr == CR_GOAL_LEN_FAST) {
1129 ext4_mb_choose_next_group_goal_fast(ac, new_cr, group);
1130 } else if (*new_cr == CR_BEST_AVAIL_LEN) {
1131 ext4_mb_choose_next_group_best_avail(ac, new_cr, group);
1134 * TODO: For CR=2, we can arrange groups in an rb tree sorted by
1135 * bb_free. But until that happens, we should never come here.
1142 * Cache the order of the largest free extent we have available in this block
1146 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
1148 struct ext4_sb_info *sbi = EXT4_SB(sb);
1151 for (i = MB_NUM_ORDERS(sb) - 1; i >= 0; i--)
1152 if (grp->bb_counters[i] > 0)
1154 /* No need to move between order lists? */
1155 if (!test_opt2(sb, MB_OPTIMIZE_SCAN) ||
1156 i == grp->bb_largest_free_order) {
1157 grp->bb_largest_free_order = i;
1161 if (grp->bb_largest_free_order >= 0) {
1162 write_lock(&sbi->s_mb_largest_free_orders_locks[
1163 grp->bb_largest_free_order]);
1164 list_del_init(&grp->bb_largest_free_order_node);
1165 write_unlock(&sbi->s_mb_largest_free_orders_locks[
1166 grp->bb_largest_free_order]);
1168 grp->bb_largest_free_order = i;
1169 if (grp->bb_largest_free_order >= 0 && grp->bb_free) {
1170 write_lock(&sbi->s_mb_largest_free_orders_locks[
1171 grp->bb_largest_free_order]);
1172 list_add_tail(&grp->bb_largest_free_order_node,
1173 &sbi->s_mb_largest_free_orders[grp->bb_largest_free_order]);
1174 write_unlock(&sbi->s_mb_largest_free_orders_locks[
1175 grp->bb_largest_free_order]);
1179 static noinline_for_stack
1180 void ext4_mb_generate_buddy(struct super_block *sb,
1181 void *buddy, void *bitmap, ext4_group_t group,
1182 struct ext4_group_info *grp)
1184 struct ext4_sb_info *sbi = EXT4_SB(sb);
1185 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
1186 ext4_grpblk_t i = 0;
1187 ext4_grpblk_t first;
1190 unsigned fragments = 0;
1191 unsigned long long period = get_cycles();
1193 /* initialize buddy from bitmap which is aggregation
1194 * of on-disk bitmap and preallocations */
1195 i = mb_find_next_zero_bit(bitmap, max, 0);
1196 grp->bb_first_free = i;
1200 i = mb_find_next_bit(bitmap, max, i);
1204 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
1206 grp->bb_counters[0]++;
1208 i = mb_find_next_zero_bit(bitmap, max, i);
1210 grp->bb_fragments = fragments;
1212 if (free != grp->bb_free) {
1213 ext4_grp_locked_error(sb, group, 0, 0,
1214 "block bitmap and bg descriptor "
1215 "inconsistent: %u vs %u free clusters",
1216 free, grp->bb_free);
1218 * If we intend to continue, we consider group descriptor
1219 * corrupt and update bb_free using bitmap value
1221 grp->bb_free = free;
1222 ext4_mark_group_bitmap_corrupted(sb, group,
1223 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1225 mb_set_largest_free_order(sb, grp);
1226 mb_update_avg_fragment_size(sb, grp);
1228 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
1230 period = get_cycles() - period;
1231 atomic_inc(&sbi->s_mb_buddies_generated);
1232 atomic64_add(period, &sbi->s_mb_generation_time);
1235 static void mb_regenerate_buddy(struct ext4_buddy *e4b)
1241 while ((buddy = mb_find_buddy(e4b, order++, &count)))
1242 mb_set_bits(buddy, 0, count);
1244 e4b->bd_info->bb_fragments = 0;
1245 memset(e4b->bd_info->bb_counters, 0,
1246 sizeof(*e4b->bd_info->bb_counters) *
1247 (e4b->bd_sb->s_blocksize_bits + 2));
1249 ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy,
1250 e4b->bd_bitmap, e4b->bd_group, e4b->bd_info);
1253 /* The buddy information is attached the buddy cache inode
1254 * for convenience. The information regarding each group
1255 * is loaded via ext4_mb_load_buddy. The information involve
1256 * block bitmap and buddy information. The information are
1257 * stored in the inode as
1260 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
1263 * one block each for bitmap and buddy information.
1264 * So for each group we take up 2 blocks. A page can
1265 * contain blocks_per_page (PAGE_SIZE / blocksize) blocks.
1266 * So it can have information regarding groups_per_page which
1267 * is blocks_per_page/2
1269 * Locking note: This routine takes the block group lock of all groups
1270 * for this page; do not hold this lock when calling this routine!
1273 static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp)
1275 ext4_group_t ngroups;
1276 unsigned int blocksize;
1277 int blocks_per_page;
1278 int groups_per_page;
1281 ext4_group_t first_group, group;
1283 struct super_block *sb;
1284 struct buffer_head *bhs;
1285 struct buffer_head **bh = NULL;
1286 struct inode *inode;
1289 struct ext4_group_info *grinfo;
1291 inode = page->mapping->host;
1293 ngroups = ext4_get_groups_count(sb);
1294 blocksize = i_blocksize(inode);
1295 blocks_per_page = PAGE_SIZE / blocksize;
1297 mb_debug(sb, "init page %lu\n", page->index);
1299 groups_per_page = blocks_per_page >> 1;
1300 if (groups_per_page == 0)
1301 groups_per_page = 1;
1303 /* allocate buffer_heads to read bitmaps */
1304 if (groups_per_page > 1) {
1305 i = sizeof(struct buffer_head *) * groups_per_page;
1306 bh = kzalloc(i, gfp);
1312 first_group = page->index * blocks_per_page / 2;
1314 /* read all groups the page covers into the cache */
1315 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
1316 if (group >= ngroups)
1319 grinfo = ext4_get_group_info(sb, group);
1323 * If page is uptodate then we came here after online resize
1324 * which added some new uninitialized group info structs, so
1325 * we must skip all initialized uptodate buddies on the page,
1326 * which may be currently in use by an allocating task.
1328 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
1332 bh[i] = ext4_read_block_bitmap_nowait(sb, group, false);
1333 if (IS_ERR(bh[i])) {
1334 err = PTR_ERR(bh[i]);
1338 mb_debug(sb, "read bitmap for group %u\n", group);
1341 /* wait for I/O completion */
1342 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
1347 err2 = ext4_wait_block_bitmap(sb, group, bh[i]);
1352 first_block = page->index * blocks_per_page;
1353 for (i = 0; i < blocks_per_page; i++) {
1354 group = (first_block + i) >> 1;
1355 if (group >= ngroups)
1358 if (!bh[group - first_group])
1359 /* skip initialized uptodate buddy */
1362 if (!buffer_verified(bh[group - first_group]))
1363 /* Skip faulty bitmaps */
1368 * data carry information regarding this
1369 * particular group in the format specified
1373 data = page_address(page) + (i * blocksize);
1374 bitmap = bh[group - first_group]->b_data;
1377 * We place the buddy block and bitmap block
1380 grinfo = ext4_get_group_info(sb, group);
1382 err = -EFSCORRUPTED;
1385 if ((first_block + i) & 1) {
1386 /* this is block of buddy */
1387 BUG_ON(incore == NULL);
1388 mb_debug(sb, "put buddy for group %u in page %lu/%x\n",
1389 group, page->index, i * blocksize);
1390 trace_ext4_mb_buddy_bitmap_load(sb, group);
1391 grinfo->bb_fragments = 0;
1392 memset(grinfo->bb_counters, 0,
1393 sizeof(*grinfo->bb_counters) *
1394 (MB_NUM_ORDERS(sb)));
1396 * incore got set to the group block bitmap below
1398 ext4_lock_group(sb, group);
1399 /* init the buddy */
1400 memset(data, 0xff, blocksize);
1401 ext4_mb_generate_buddy(sb, data, incore, group, grinfo);
1402 ext4_unlock_group(sb, group);
1405 /* this is block of bitmap */
1406 BUG_ON(incore != NULL);
1407 mb_debug(sb, "put bitmap for group %u in page %lu/%x\n",
1408 group, page->index, i * blocksize);
1409 trace_ext4_mb_bitmap_load(sb, group);
1411 /* see comments in ext4_mb_put_pa() */
1412 ext4_lock_group(sb, group);
1413 memcpy(data, bitmap, blocksize);
1415 /* mark all preallocated blks used in in-core bitmap */
1416 ext4_mb_generate_from_pa(sb, data, group);
1417 WARN_ON_ONCE(!RB_EMPTY_ROOT(&grinfo->bb_free_root));
1418 ext4_unlock_group(sb, group);
1420 /* set incore so that the buddy information can be
1421 * generated using this
1426 SetPageUptodate(page);
1430 for (i = 0; i < groups_per_page; i++)
1439 * Lock the buddy and bitmap pages. This make sure other parallel init_group
1440 * on the same buddy page doesn't happen whild holding the buddy page lock.
1441 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
1442 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
1444 static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
1445 ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp)
1447 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
1448 int block, pnum, poff;
1449 int blocks_per_page;
1452 e4b->bd_buddy_page = NULL;
1453 e4b->bd_bitmap_page = NULL;
1455 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1457 * the buddy cache inode stores the block bitmap
1458 * and buddy information in consecutive blocks.
1459 * So for each group we need two blocks.
1462 pnum = block / blocks_per_page;
1463 poff = block % blocks_per_page;
1464 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1467 BUG_ON(page->mapping != inode->i_mapping);
1468 e4b->bd_bitmap_page = page;
1469 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1471 if (blocks_per_page >= 2) {
1472 /* buddy and bitmap are on the same page */
1476 /* blocks_per_page == 1, hence we need another page for the buddy */
1477 page = find_or_create_page(inode->i_mapping, block + 1, gfp);
1480 BUG_ON(page->mapping != inode->i_mapping);
1481 e4b->bd_buddy_page = page;
1485 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1487 if (e4b->bd_bitmap_page) {
1488 unlock_page(e4b->bd_bitmap_page);
1489 put_page(e4b->bd_bitmap_page);
1491 if (e4b->bd_buddy_page) {
1492 unlock_page(e4b->bd_buddy_page);
1493 put_page(e4b->bd_buddy_page);
1498 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1499 * block group lock of all groups for this page; do not hold the BG lock when
1500 * calling this routine!
1502 static noinline_for_stack
1503 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp)
1506 struct ext4_group_info *this_grp;
1507 struct ext4_buddy e4b;
1512 mb_debug(sb, "init group %u\n", group);
1513 this_grp = ext4_get_group_info(sb, group);
1515 return -EFSCORRUPTED;
1518 * This ensures that we don't reinit the buddy cache
1519 * page which map to the group from which we are already
1520 * allocating. If we are looking at the buddy cache we would
1521 * have taken a reference using ext4_mb_load_buddy and that
1522 * would have pinned buddy page to page cache.
1523 * The call to ext4_mb_get_buddy_page_lock will mark the
1526 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp);
1527 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1529 * somebody initialized the group
1530 * return without doing anything
1535 page = e4b.bd_bitmap_page;
1536 ret = ext4_mb_init_cache(page, NULL, gfp);
1539 if (!PageUptodate(page)) {
1544 if (e4b.bd_buddy_page == NULL) {
1546 * If both the bitmap and buddy are in
1547 * the same page we don't need to force
1553 /* init buddy cache */
1554 page = e4b.bd_buddy_page;
1555 ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp);
1558 if (!PageUptodate(page)) {
1563 ext4_mb_put_buddy_page_lock(&e4b);
1568 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1569 * block group lock of all groups for this page; do not hold the BG lock when
1570 * calling this routine!
1572 static noinline_for_stack int
1573 ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group,
1574 struct ext4_buddy *e4b, gfp_t gfp)
1576 int blocks_per_page;
1582 struct ext4_group_info *grp;
1583 struct ext4_sb_info *sbi = EXT4_SB(sb);
1584 struct inode *inode = sbi->s_buddy_cache;
1587 mb_debug(sb, "load group %u\n", group);
1589 blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1590 grp = ext4_get_group_info(sb, group);
1592 return -EFSCORRUPTED;
1594 e4b->bd_blkbits = sb->s_blocksize_bits;
1597 e4b->bd_group = group;
1598 e4b->bd_buddy_page = NULL;
1599 e4b->bd_bitmap_page = NULL;
1601 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1603 * we need full data about the group
1604 * to make a good selection
1606 ret = ext4_mb_init_group(sb, group, gfp);
1612 * the buddy cache inode stores the block bitmap
1613 * and buddy information in consecutive blocks.
1614 * So for each group we need two blocks.
1617 pnum = block / blocks_per_page;
1618 poff = block % blocks_per_page;
1620 /* we could use find_or_create_page(), but it locks page
1621 * what we'd like to avoid in fast path ... */
1622 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1623 if (page == NULL || !PageUptodate(page)) {
1626 * drop the page reference and try
1627 * to get the page with lock. If we
1628 * are not uptodate that implies
1629 * somebody just created the page but
1630 * is yet to initialize the same. So
1631 * wait for it to initialize.
1634 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1636 if (WARN_RATELIMIT(page->mapping != inode->i_mapping,
1637 "ext4: bitmap's paging->mapping != inode->i_mapping\n")) {
1638 /* should never happen */
1643 if (!PageUptodate(page)) {
1644 ret = ext4_mb_init_cache(page, NULL, gfp);
1649 mb_cmp_bitmaps(e4b, page_address(page) +
1650 (poff * sb->s_blocksize));
1659 if (!PageUptodate(page)) {
1664 /* Pages marked accessed already */
1665 e4b->bd_bitmap_page = page;
1666 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1669 pnum = block / blocks_per_page;
1670 poff = block % blocks_per_page;
1672 page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1673 if (page == NULL || !PageUptodate(page)) {
1676 page = find_or_create_page(inode->i_mapping, pnum, gfp);
1678 if (WARN_RATELIMIT(page->mapping != inode->i_mapping,
1679 "ext4: buddy bitmap's page->mapping != inode->i_mapping\n")) {
1680 /* should never happen */
1685 if (!PageUptodate(page)) {
1686 ret = ext4_mb_init_cache(page, e4b->bd_bitmap,
1700 if (!PageUptodate(page)) {
1705 /* Pages marked accessed already */
1706 e4b->bd_buddy_page = page;
1707 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1714 if (e4b->bd_bitmap_page)
1715 put_page(e4b->bd_bitmap_page);
1717 e4b->bd_buddy = NULL;
1718 e4b->bd_bitmap = NULL;
1722 static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1723 struct ext4_buddy *e4b)
1725 return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS);
1728 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1730 if (e4b->bd_bitmap_page)
1731 put_page(e4b->bd_bitmap_page);
1732 if (e4b->bd_buddy_page)
1733 put_page(e4b->bd_buddy_page);
1737 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1742 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1743 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1745 while (order <= e4b->bd_blkbits + 1) {
1746 bb = mb_find_buddy(e4b, order, &max);
1747 if (!mb_test_bit(block >> order, bb)) {
1748 /* this block is part of buddy of order 'order' */
1756 static void mb_clear_bits(void *bm, int cur, int len)
1762 if ((cur & 31) == 0 && (len - cur) >= 32) {
1763 /* fast path: clear whole word at once */
1764 addr = bm + (cur >> 3);
1769 mb_clear_bit(cur, bm);
1774 /* clear bits in given range
1775 * will return first found zero bit if any, -1 otherwise
1777 static int mb_test_and_clear_bits(void *bm, int cur, int len)
1784 if ((cur & 31) == 0 && (len - cur) >= 32) {
1785 /* fast path: clear whole word at once */
1786 addr = bm + (cur >> 3);
1787 if (*addr != (__u32)(-1) && zero_bit == -1)
1788 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1793 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1801 void mb_set_bits(void *bm, int cur, int len)
1807 if ((cur & 31) == 0 && (len - cur) >= 32) {
1808 /* fast path: set whole word at once */
1809 addr = bm + (cur >> 3);
1814 mb_set_bit(cur, bm);
1819 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1821 if (mb_test_bit(*bit + side, bitmap)) {
1822 mb_clear_bit(*bit, bitmap);
1828 mb_set_bit(*bit, bitmap);
1833 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1837 void *buddy = mb_find_buddy(e4b, order, &max);
1842 /* Bits in range [first; last] are known to be set since
1843 * corresponding blocks were allocated. Bits in range
1844 * (first; last) will stay set because they form buddies on
1845 * upper layer. We just deal with borders if they don't
1846 * align with upper layer and then go up.
1847 * Releasing entire group is all about clearing
1848 * single bit of highest order buddy.
1852 * ---------------------------------
1854 * ---------------------------------
1855 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1856 * ---------------------------------
1858 * \_____________________/
1860 * Neither [1] nor [6] is aligned to above layer.
1861 * Left neighbour [0] is free, so mark it busy,
1862 * decrease bb_counters and extend range to
1864 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1865 * mark [6] free, increase bb_counters and shrink range to
1867 * Then shift range to [0; 2], go up and do the same.
1872 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1874 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1879 buddy2 = mb_find_buddy(e4b, order, &max);
1881 mb_clear_bits(buddy, first, last - first + 1);
1882 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1891 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1892 int first, int count)
1894 int left_is_free = 0;
1895 int right_is_free = 0;
1897 int last = first + count - 1;
1898 struct super_block *sb = e4b->bd_sb;
1900 if (WARN_ON(count == 0))
1902 BUG_ON(last >= (sb->s_blocksize << 3));
1903 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1904 /* Don't bother if the block group is corrupt. */
1905 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1908 mb_check_buddy(e4b);
1909 mb_free_blocks_double(inode, e4b, first, count);
1911 /* access memory sequentially: check left neighbour,
1912 * clear range and then check right neighbour
1915 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1916 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1917 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1918 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1920 if (unlikely(block != -1)) {
1921 struct ext4_sb_info *sbi = EXT4_SB(sb);
1922 ext4_fsblk_t blocknr;
1925 * Fastcommit replay can free already freed blocks which
1926 * corrupts allocation info. Regenerate it.
1928 if (sbi->s_mount_state & EXT4_FC_REPLAY) {
1929 mb_regenerate_buddy(e4b);
1933 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1934 blocknr += EXT4_C2B(sbi, block);
1935 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
1936 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
1937 ext4_grp_locked_error(sb, e4b->bd_group,
1938 inode ? inode->i_ino : 0, blocknr,
1939 "freeing already freed block (bit %u); block bitmap corrupt.",
1944 this_cpu_inc(discard_pa_seq);
1945 e4b->bd_info->bb_free += count;
1946 if (first < e4b->bd_info->bb_first_free)
1947 e4b->bd_info->bb_first_free = first;
1949 /* let's maintain fragments counter */
1950 if (left_is_free && right_is_free)
1951 e4b->bd_info->bb_fragments--;
1952 else if (!left_is_free && !right_is_free)
1953 e4b->bd_info->bb_fragments++;
1955 /* buddy[0] == bd_bitmap is a special case, so handle
1956 * it right away and let mb_buddy_mark_free stay free of
1957 * zero order checks.
1958 * Check if neighbours are to be coaleasced,
1959 * adjust bitmap bb_counters and borders appropriately.
1962 first += !left_is_free;
1963 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1966 last -= !right_is_free;
1967 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1971 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1973 mb_set_largest_free_order(sb, e4b->bd_info);
1974 mb_update_avg_fragment_size(sb, e4b->bd_info);
1976 mb_check_buddy(e4b);
1979 static int mb_find_extent(struct ext4_buddy *e4b, int block,
1980 int needed, struct ext4_free_extent *ex)
1982 int max, order, next;
1985 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1988 buddy = mb_find_buddy(e4b, 0, &max);
1989 BUG_ON(buddy == NULL);
1990 BUG_ON(block >= max);
1991 if (mb_test_bit(block, buddy)) {
1998 /* find actual order */
1999 order = mb_find_order_for_block(e4b, block);
2001 ex->fe_len = (1 << order) - (block & ((1 << order) - 1));
2002 ex->fe_start = block;
2003 ex->fe_group = e4b->bd_group;
2005 block = block >> order;
2007 while (needed > ex->fe_len &&
2008 mb_find_buddy(e4b, order, &max)) {
2010 if (block + 1 >= max)
2013 next = (block + 1) * (1 << order);
2014 if (mb_test_bit(next, e4b->bd_bitmap))
2017 order = mb_find_order_for_block(e4b, next);
2019 block = next >> order;
2020 ex->fe_len += 1 << order;
2023 if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) {
2024 /* Should never happen! (but apparently sometimes does?!?) */
2026 ext4_grp_locked_error(e4b->bd_sb, e4b->bd_group, 0, 0,
2027 "corruption or bug in mb_find_extent "
2028 "block=%d, order=%d needed=%d ex=%u/%d/%d@%u",
2029 block, order, needed, ex->fe_group, ex->fe_start,
2030 ex->fe_len, ex->fe_logical);
2038 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
2044 int start = ex->fe_start;
2045 int len = ex->fe_len;
2051 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
2052 BUG_ON(e4b->bd_group != ex->fe_group);
2053 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
2054 mb_check_buddy(e4b);
2055 mb_mark_used_double(e4b, start, len);
2057 this_cpu_inc(discard_pa_seq);
2058 e4b->bd_info->bb_free -= len;
2059 if (e4b->bd_info->bb_first_free == start)
2060 e4b->bd_info->bb_first_free += len;
2062 /* let's maintain fragments counter */
2064 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
2065 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
2066 max = !mb_test_bit(start + len, e4b->bd_bitmap);
2068 e4b->bd_info->bb_fragments++;
2069 else if (!mlen && !max)
2070 e4b->bd_info->bb_fragments--;
2072 /* let's maintain buddy itself */
2075 ord = mb_find_order_for_block(e4b, start);
2077 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
2078 /* the whole chunk may be allocated at once! */
2081 buddy = mb_find_buddy(e4b, ord, &max);
2084 BUG_ON((start >> ord) >= max);
2085 mb_set_bit(start >> ord, buddy);
2086 e4b->bd_info->bb_counters[ord]--;
2093 /* store for history */
2095 ret = len | (ord << 16);
2097 /* we have to split large buddy */
2099 buddy = mb_find_buddy(e4b, ord, &max);
2100 mb_set_bit(start >> ord, buddy);
2101 e4b->bd_info->bb_counters[ord]--;
2104 cur = (start >> ord) & ~1U;
2105 buddy = mb_find_buddy(e4b, ord, &max);
2106 mb_clear_bit(cur, buddy);
2107 mb_clear_bit(cur + 1, buddy);
2108 e4b->bd_info->bb_counters[ord]++;
2109 e4b->bd_info->bb_counters[ord]++;
2112 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
2114 mb_update_avg_fragment_size(e4b->bd_sb, e4b->bd_info);
2115 mb_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
2116 mb_check_buddy(e4b);
2122 * Must be called under group lock!
2124 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
2125 struct ext4_buddy *e4b)
2127 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2130 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
2131 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2133 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
2134 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
2135 ret = mb_mark_used(e4b, &ac->ac_b_ex);
2137 /* preallocation can change ac_b_ex, thus we store actually
2138 * allocated blocks for history */
2139 ac->ac_f_ex = ac->ac_b_ex;
2141 ac->ac_status = AC_STATUS_FOUND;
2142 ac->ac_tail = ret & 0xffff;
2143 ac->ac_buddy = ret >> 16;
2146 * take the page reference. We want the page to be pinned
2147 * so that we don't get a ext4_mb_init_cache_call for this
2148 * group until we update the bitmap. That would mean we
2149 * double allocate blocks. The reference is dropped
2150 * in ext4_mb_release_context
2152 ac->ac_bitmap_page = e4b->bd_bitmap_page;
2153 get_page(ac->ac_bitmap_page);
2154 ac->ac_buddy_page = e4b->bd_buddy_page;
2155 get_page(ac->ac_buddy_page);
2156 /* store last allocated for subsequent stream allocation */
2157 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2158 spin_lock(&sbi->s_md_lock);
2159 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
2160 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
2161 spin_unlock(&sbi->s_md_lock);
2164 * As we've just preallocated more space than
2165 * user requested originally, we store allocated
2166 * space in a special descriptor.
2168 if (ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
2169 ext4_mb_new_preallocation(ac);
2173 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
2174 struct ext4_buddy *e4b,
2177 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2178 struct ext4_free_extent *bex = &ac->ac_b_ex;
2179 struct ext4_free_extent *gex = &ac->ac_g_ex;
2181 if (ac->ac_status == AC_STATUS_FOUND)
2184 * We don't want to scan for a whole year
2186 if (ac->ac_found > sbi->s_mb_max_to_scan &&
2187 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2188 ac->ac_status = AC_STATUS_BREAK;
2193 * Haven't found good chunk so far, let's continue
2195 if (bex->fe_len < gex->fe_len)
2198 if (finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
2199 ext4_mb_use_best_found(ac, e4b);
2203 * The routine checks whether found extent is good enough. If it is,
2204 * then the extent gets marked used and flag is set to the context
2205 * to stop scanning. Otherwise, the extent is compared with the
2206 * previous found extent and if new one is better, then it's stored
2207 * in the context. Later, the best found extent will be used, if
2208 * mballoc can't find good enough extent.
2210 * The algorithm used is roughly as follows:
2212 * * If free extent found is exactly as big as goal, then
2213 * stop the scan and use it immediately
2215 * * If free extent found is smaller than goal, then keep retrying
2216 * upto a max of sbi->s_mb_max_to_scan times (default 200). After
2217 * that stop scanning and use whatever we have.
2219 * * If free extent found is bigger than goal, then keep retrying
2220 * upto a max of sbi->s_mb_min_to_scan times (default 10) before
2221 * stopping the scan and using the extent.
2224 * FIXME: real allocation policy is to be designed yet!
2226 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
2227 struct ext4_free_extent *ex,
2228 struct ext4_buddy *e4b)
2230 struct ext4_free_extent *bex = &ac->ac_b_ex;
2231 struct ext4_free_extent *gex = &ac->ac_g_ex;
2233 BUG_ON(ex->fe_len <= 0);
2234 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
2235 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
2236 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
2239 ac->ac_cX_found[ac->ac_criteria]++;
2242 * The special case - take what you catch first
2244 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2246 ext4_mb_use_best_found(ac, e4b);
2251 * Let's check whether the chuck is good enough
2253 if (ex->fe_len == gex->fe_len) {
2255 ext4_mb_use_best_found(ac, e4b);
2260 * If this is first found extent, just store it in the context
2262 if (bex->fe_len == 0) {
2268 * If new found extent is better, store it in the context
2270 if (bex->fe_len < gex->fe_len) {
2271 /* if the request isn't satisfied, any found extent
2272 * larger than previous best one is better */
2273 if (ex->fe_len > bex->fe_len)
2275 } else if (ex->fe_len > gex->fe_len) {
2276 /* if the request is satisfied, then we try to find
2277 * an extent that still satisfy the request, but is
2278 * smaller than previous one */
2279 if (ex->fe_len < bex->fe_len)
2283 ext4_mb_check_limits(ac, e4b, 0);
2286 static noinline_for_stack
2287 void ext4_mb_try_best_found(struct ext4_allocation_context *ac,
2288 struct ext4_buddy *e4b)
2290 struct ext4_free_extent ex = ac->ac_b_ex;
2291 ext4_group_t group = ex.fe_group;
2295 BUG_ON(ex.fe_len <= 0);
2296 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
2300 ext4_lock_group(ac->ac_sb, group);
2301 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
2304 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
2308 ext4_mb_use_best_found(ac, e4b);
2312 ext4_unlock_group(ac->ac_sb, group);
2313 ext4_mb_unload_buddy(e4b);
2316 static noinline_for_stack
2317 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
2318 struct ext4_buddy *e4b)
2320 ext4_group_t group = ac->ac_g_ex.fe_group;
2323 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2324 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2325 struct ext4_free_extent ex;
2328 return -EFSCORRUPTED;
2329 if (!(ac->ac_flags & (EXT4_MB_HINT_TRY_GOAL | EXT4_MB_HINT_GOAL_ONLY)))
2331 if (grp->bb_free == 0)
2334 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
2338 ext4_lock_group(ac->ac_sb, group);
2339 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
2342 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
2343 ac->ac_g_ex.fe_len, &ex);
2344 ex.fe_logical = 0xDEADFA11; /* debug value */
2346 if (max >= ac->ac_g_ex.fe_len &&
2347 ac->ac_g_ex.fe_len == EXT4_B2C(sbi, sbi->s_stripe)) {
2350 start = ext4_grp_offs_to_block(ac->ac_sb, &ex);
2351 /* use do_div to get remainder (would be 64-bit modulo) */
2352 if (do_div(start, sbi->s_stripe) == 0) {
2355 ext4_mb_use_best_found(ac, e4b);
2357 } else if (max >= ac->ac_g_ex.fe_len) {
2358 BUG_ON(ex.fe_len <= 0);
2359 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
2360 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
2363 ext4_mb_use_best_found(ac, e4b);
2364 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
2365 /* Sometimes, caller may want to merge even small
2366 * number of blocks to an existing extent */
2367 BUG_ON(ex.fe_len <= 0);
2368 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
2369 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
2372 ext4_mb_use_best_found(ac, e4b);
2375 ext4_unlock_group(ac->ac_sb, group);
2376 ext4_mb_unload_buddy(e4b);
2382 * The routine scans buddy structures (not bitmap!) from given order
2383 * to max order and tries to find big enough chunk to satisfy the req
2385 static noinline_for_stack
2386 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
2387 struct ext4_buddy *e4b)
2389 struct super_block *sb = ac->ac_sb;
2390 struct ext4_group_info *grp = e4b->bd_info;
2396 BUG_ON(ac->ac_2order <= 0);
2397 for (i = ac->ac_2order; i < MB_NUM_ORDERS(sb); i++) {
2398 if (grp->bb_counters[i] == 0)
2401 buddy = mb_find_buddy(e4b, i, &max);
2402 if (WARN_RATELIMIT(buddy == NULL,
2403 "ext4: mb_simple_scan_group: mb_find_buddy failed, (%d)\n", i))
2406 k = mb_find_next_zero_bit(buddy, max, 0);
2408 ext4_mark_group_bitmap_corrupted(ac->ac_sb,
2410 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2411 ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0,
2412 "%d free clusters of order %d. But found 0",
2413 grp->bb_counters[i], i);
2417 ac->ac_cX_found[ac->ac_criteria]++;
2419 ac->ac_b_ex.fe_len = 1 << i;
2420 ac->ac_b_ex.fe_start = k << i;
2421 ac->ac_b_ex.fe_group = e4b->bd_group;
2423 ext4_mb_use_best_found(ac, e4b);
2425 BUG_ON(ac->ac_f_ex.fe_len != ac->ac_g_ex.fe_len);
2427 if (EXT4_SB(sb)->s_mb_stats)
2428 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
2435 * The routine scans the group and measures all found extents.
2436 * In order to optimize scanning, caller must pass number of
2437 * free blocks in the group, so the routine can know upper limit.
2439 static noinline_for_stack
2440 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
2441 struct ext4_buddy *e4b)
2443 struct super_block *sb = ac->ac_sb;
2444 void *bitmap = e4b->bd_bitmap;
2445 struct ext4_free_extent ex;
2449 free = e4b->bd_info->bb_free;
2450 if (WARN_ON(free <= 0))
2453 i = e4b->bd_info->bb_first_free;
2455 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
2456 i = mb_find_next_zero_bit(bitmap,
2457 EXT4_CLUSTERS_PER_GROUP(sb), i);
2458 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
2460 * IF we have corrupt bitmap, we won't find any
2461 * free blocks even though group info says we
2464 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2465 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2466 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2467 "%d free clusters as per "
2468 "group info. But bitmap says 0",
2473 if (!ext4_mb_cr_expensive(ac->ac_criteria)) {
2475 * In CR_GOAL_LEN_FAST and CR_BEST_AVAIL_LEN, we are
2476 * sure that this group will have a large enough
2477 * continuous free extent, so skip over the smaller free
2480 j = mb_find_next_bit(bitmap,
2481 EXT4_CLUSTERS_PER_GROUP(sb), i);
2484 if (freelen < ac->ac_g_ex.fe_len) {
2491 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
2492 if (WARN_ON(ex.fe_len <= 0))
2494 if (free < ex.fe_len) {
2495 ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group,
2496 EXT4_GROUP_INFO_BBITMAP_CORRUPT);
2497 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
2498 "%d free clusters as per "
2499 "group info. But got %d blocks",
2502 * The number of free blocks differs. This mostly
2503 * indicate that the bitmap is corrupt. So exit
2504 * without claiming the space.
2508 ex.fe_logical = 0xDEADC0DE; /* debug value */
2509 ext4_mb_measure_extent(ac, &ex, e4b);
2515 ext4_mb_check_limits(ac, e4b, 1);
2519 * This is a special case for storages like raid5
2520 * we try to find stripe-aligned chunks for stripe-size-multiple requests
2522 static noinline_for_stack
2523 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
2524 struct ext4_buddy *e4b)
2526 struct super_block *sb = ac->ac_sb;
2527 struct ext4_sb_info *sbi = EXT4_SB(sb);
2528 void *bitmap = e4b->bd_bitmap;
2529 struct ext4_free_extent ex;
2530 ext4_fsblk_t first_group_block;
2532 ext4_grpblk_t i, stripe;
2535 BUG_ON(sbi->s_stripe == 0);
2537 /* find first stripe-aligned block in group */
2538 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
2540 a = first_group_block + sbi->s_stripe - 1;
2541 do_div(a, sbi->s_stripe);
2542 i = (a * sbi->s_stripe) - first_group_block;
2544 stripe = EXT4_B2C(sbi, sbi->s_stripe);
2545 i = EXT4_B2C(sbi, i);
2546 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
2547 if (!mb_test_bit(i, bitmap)) {
2548 max = mb_find_extent(e4b, i, stripe, &ex);
2549 if (max >= stripe) {
2551 ac->ac_cX_found[ac->ac_criteria]++;
2552 ex.fe_logical = 0xDEADF00D; /* debug value */
2554 ext4_mb_use_best_found(ac, e4b);
2563 * This is also called BEFORE we load the buddy bitmap.
2564 * Returns either 1 or 0 indicating that the group is either suitable
2565 * for the allocation or not.
2567 static bool ext4_mb_good_group(struct ext4_allocation_context *ac,
2568 ext4_group_t group, enum criteria cr)
2570 ext4_grpblk_t free, fragments;
2571 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2572 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2574 BUG_ON(cr < CR_POWER2_ALIGNED || cr >= EXT4_MB_NUM_CRS);
2576 if (unlikely(!grp || EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2579 free = grp->bb_free;
2583 fragments = grp->bb_fragments;
2588 case CR_POWER2_ALIGNED:
2589 BUG_ON(ac->ac_2order == 0);
2591 /* Avoid using the first bg of a flexgroup for data files */
2592 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2593 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2594 ((group % flex_size) == 0))
2597 if (free < ac->ac_g_ex.fe_len)
2600 if (ac->ac_2order >= MB_NUM_ORDERS(ac->ac_sb))
2603 if (grp->bb_largest_free_order < ac->ac_2order)
2607 case CR_GOAL_LEN_FAST:
2608 case CR_BEST_AVAIL_LEN:
2609 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2612 case CR_GOAL_LEN_SLOW:
2613 if (free >= ac->ac_g_ex.fe_len)
2626 * This could return negative error code if something goes wrong
2627 * during ext4_mb_init_group(). This should not be called with
2628 * ext4_lock_group() held.
2630 * Note: because we are conditionally operating with the group lock in
2631 * the EXT4_MB_STRICT_CHECK case, we need to fake out sparse in this
2632 * function using __acquire and __release. This means we need to be
2633 * super careful before messing with the error path handling via "goto
2636 static int ext4_mb_good_group_nolock(struct ext4_allocation_context *ac,
2637 ext4_group_t group, enum criteria cr)
2639 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2640 struct super_block *sb = ac->ac_sb;
2641 struct ext4_sb_info *sbi = EXT4_SB(sb);
2642 bool should_lock = ac->ac_flags & EXT4_MB_STRICT_CHECK;
2647 return -EFSCORRUPTED;
2648 if (sbi->s_mb_stats)
2649 atomic64_inc(&sbi->s_bal_cX_groups_considered[ac->ac_criteria]);
2651 ext4_lock_group(sb, group);
2652 __release(ext4_group_lock_ptr(sb, group));
2654 free = grp->bb_free;
2658 * In all criterias except CR_ANY_FREE we try to avoid groups that
2659 * can't possibly satisfy the full goal request due to insufficient
2662 if (cr < CR_ANY_FREE && free < ac->ac_g_ex.fe_len)
2664 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2667 __acquire(ext4_group_lock_ptr(sb, group));
2668 ext4_unlock_group(sb, group);
2671 /* We only do this if the grp has never been initialized */
2672 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2673 struct ext4_group_desc *gdp =
2674 ext4_get_group_desc(sb, group, NULL);
2678 * cr=CR_POWER2_ALIGNED/CR_GOAL_LEN_FAST is a very optimistic
2679 * search to find large good chunks almost for free. If buddy
2680 * data is not ready, then this optimization makes no sense. But
2681 * we never skip the first block group in a flex_bg, since this
2682 * gets used for metadata block allocation, and we want to make
2683 * sure we locate metadata blocks in the first block group in
2684 * the flex_bg if possible.
2686 if (!ext4_mb_cr_expensive(cr) &&
2687 (!sbi->s_log_groups_per_flex ||
2688 ((group & ((1 << sbi->s_log_groups_per_flex) - 1)) != 0)) &&
2689 !(ext4_has_group_desc_csum(sb) &&
2690 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))))
2692 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
2698 ext4_lock_group(sb, group);
2699 __release(ext4_group_lock_ptr(sb, group));
2701 ret = ext4_mb_good_group(ac, group, cr);
2704 __acquire(ext4_group_lock_ptr(sb, group));
2705 ext4_unlock_group(sb, group);
2711 * Start prefetching @nr block bitmaps starting at @group.
2712 * Return the next group which needs to be prefetched.
2714 ext4_group_t ext4_mb_prefetch(struct super_block *sb, ext4_group_t group,
2715 unsigned int nr, int *cnt)
2717 ext4_group_t ngroups = ext4_get_groups_count(sb);
2718 struct buffer_head *bh;
2719 struct blk_plug plug;
2721 blk_start_plug(&plug);
2723 struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
2725 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
2728 * Prefetch block groups with free blocks; but don't
2729 * bother if it is marked uninitialized on disk, since
2730 * it won't require I/O to read. Also only try to
2731 * prefetch once, so we avoid getblk() call, which can
2734 if (gdp && grp && !EXT4_MB_GRP_TEST_AND_SET_READ(grp) &&
2735 EXT4_MB_GRP_NEED_INIT(grp) &&
2736 ext4_free_group_clusters(sb, gdp) > 0 ) {
2737 bh = ext4_read_block_bitmap_nowait(sb, group, true);
2738 if (bh && !IS_ERR(bh)) {
2739 if (!buffer_uptodate(bh) && cnt)
2744 if (++group >= ngroups)
2747 blk_finish_plug(&plug);
2752 * Prefetching reads the block bitmap into the buffer cache; but we
2753 * need to make sure that the buddy bitmap in the page cache has been
2754 * initialized. Note that ext4_mb_init_group() will block if the I/O
2755 * is not yet completed, or indeed if it was not initiated by
2756 * ext4_mb_prefetch did not start the I/O.
2758 * TODO: We should actually kick off the buddy bitmap setup in a work
2759 * queue when the buffer I/O is completed, so that we don't block
2760 * waiting for the block allocation bitmap read to finish when
2761 * ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator().
2763 void ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group,
2766 struct ext4_group_desc *gdp;
2767 struct ext4_group_info *grp;
2771 group = ext4_get_groups_count(sb);
2773 gdp = ext4_get_group_desc(sb, group, NULL);
2774 grp = ext4_get_group_info(sb, group);
2776 if (grp && gdp && EXT4_MB_GRP_NEED_INIT(grp) &&
2777 ext4_free_group_clusters(sb, gdp) > 0) {
2778 if (ext4_mb_init_group(sb, group, GFP_NOFS))
2784 static noinline_for_stack int
2785 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2787 ext4_group_t prefetch_grp = 0, ngroups, group, i;
2788 enum criteria new_cr, cr = CR_GOAL_LEN_FAST;
2789 int err = 0, first_err = 0;
2790 unsigned int nr = 0, prefetch_ios = 0;
2791 struct ext4_sb_info *sbi;
2792 struct super_block *sb;
2793 struct ext4_buddy e4b;
2798 ngroups = ext4_get_groups_count(sb);
2799 /* non-extent files are limited to low blocks/groups */
2800 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2801 ngroups = sbi->s_blockfile_groups;
2803 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2805 /* first, try the goal */
2806 err = ext4_mb_find_by_goal(ac, &e4b);
2807 if (err || ac->ac_status == AC_STATUS_FOUND)
2810 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2814 * ac->ac_2order is set only if the fe_len is a power of 2
2815 * if ac->ac_2order is set we also set criteria to CR_POWER2_ALIGNED
2816 * so that we try exact allocation using buddy.
2818 i = fls(ac->ac_g_ex.fe_len);
2821 * We search using buddy data only if the order of the request
2822 * is greater than equal to the sbi_s_mb_order2_reqs
2823 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2824 * We also support searching for power-of-two requests only for
2825 * requests upto maximum buddy size we have constructed.
2827 if (i >= sbi->s_mb_order2_reqs && i <= MB_NUM_ORDERS(sb)) {
2828 if (is_power_of_2(ac->ac_g_ex.fe_len))
2829 ac->ac_2order = array_index_nospec(i - 1,
2833 /* if stream allocation is enabled, use global goal */
2834 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2835 /* TBD: may be hot point */
2836 spin_lock(&sbi->s_md_lock);
2837 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2838 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2839 spin_unlock(&sbi->s_md_lock);
2843 * Let's just scan groups to find more-less suitable blocks We
2844 * start with CR_GOAL_LEN_FAST, unless it is power of 2
2845 * aligned, in which case let's do that faster approach first.
2848 cr = CR_POWER2_ALIGNED;
2850 for (; cr < EXT4_MB_NUM_CRS && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2851 ac->ac_criteria = cr;
2853 * searching for the right group start
2854 * from the goal value specified
2856 group = ac->ac_g_ex.fe_group;
2857 ac->ac_groups_linear_remaining = sbi->s_mb_max_linear_groups;
2858 prefetch_grp = group;
2860 for (i = 0, new_cr = cr; i < ngroups; i++,
2861 ext4_mb_choose_next_group(ac, &new_cr, &group, ngroups)) {
2871 * Batch reads of the block allocation bitmaps
2872 * to get multiple READs in flight; limit
2873 * prefetching at inexpensive CR, otherwise mballoc
2874 * can spend a lot of time loading imperfect groups
2876 if ((prefetch_grp == group) &&
2877 (ext4_mb_cr_expensive(cr) ||
2878 prefetch_ios < sbi->s_mb_prefetch_limit)) {
2879 nr = sbi->s_mb_prefetch;
2880 if (ext4_has_feature_flex_bg(sb)) {
2881 nr = 1 << sbi->s_log_groups_per_flex;
2882 nr -= group & (nr - 1);
2883 nr = min(nr, sbi->s_mb_prefetch);
2885 prefetch_grp = ext4_mb_prefetch(sb, group,
2889 /* This now checks without needing the buddy page */
2890 ret = ext4_mb_good_group_nolock(ac, group, cr);
2897 err = ext4_mb_load_buddy(sb, group, &e4b);
2901 ext4_lock_group(sb, group);
2904 * We need to check again after locking the
2907 ret = ext4_mb_good_group(ac, group, cr);
2909 ext4_unlock_group(sb, group);
2910 ext4_mb_unload_buddy(&e4b);
2914 ac->ac_groups_scanned++;
2915 if (cr == CR_POWER2_ALIGNED)
2916 ext4_mb_simple_scan_group(ac, &e4b);
2918 bool is_stripe_aligned = sbi->s_stripe &&
2919 !(ac->ac_g_ex.fe_len %
2920 EXT4_B2C(sbi, sbi->s_stripe));
2922 if ((cr == CR_GOAL_LEN_FAST ||
2923 cr == CR_BEST_AVAIL_LEN) &&
2925 ext4_mb_scan_aligned(ac, &e4b);
2927 if (ac->ac_status == AC_STATUS_CONTINUE)
2928 ext4_mb_complex_scan_group(ac, &e4b);
2931 ext4_unlock_group(sb, group);
2932 ext4_mb_unload_buddy(&e4b);
2934 if (ac->ac_status != AC_STATUS_CONTINUE)
2937 /* Processed all groups and haven't found blocks */
2938 if (sbi->s_mb_stats && i == ngroups)
2939 atomic64_inc(&sbi->s_bal_cX_failed[cr]);
2941 if (i == ngroups && ac->ac_criteria == CR_BEST_AVAIL_LEN)
2942 /* Reset goal length to original goal length before
2943 * falling into CR_GOAL_LEN_SLOW */
2944 ac->ac_g_ex.fe_len = ac->ac_orig_goal_len;
2947 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2948 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2950 * We've been searching too long. Let's try to allocate
2951 * the best chunk we've found so far
2953 ext4_mb_try_best_found(ac, &e4b);
2954 if (ac->ac_status != AC_STATUS_FOUND) {
2956 * Someone more lucky has already allocated it.
2957 * The only thing we can do is just take first
2960 lost = atomic_inc_return(&sbi->s_mb_lost_chunks);
2961 mb_debug(sb, "lost chunk, group: %u, start: %d, len: %d, lost: %d\n",
2962 ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start,
2963 ac->ac_b_ex.fe_len, lost);
2965 ac->ac_b_ex.fe_group = 0;
2966 ac->ac_b_ex.fe_start = 0;
2967 ac->ac_b_ex.fe_len = 0;
2968 ac->ac_status = AC_STATUS_CONTINUE;
2969 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2975 if (sbi->s_mb_stats && ac->ac_status == AC_STATUS_FOUND)
2976 atomic64_inc(&sbi->s_bal_cX_hits[ac->ac_criteria]);
2978 if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2981 mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n",
2982 ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status,
2983 ac->ac_flags, cr, err);
2986 ext4_mb_prefetch_fini(sb, prefetch_grp, nr);
2991 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2993 struct super_block *sb = pde_data(file_inode(seq->file));
2996 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2999 return (void *) ((unsigned long) group);
3002 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
3004 struct super_block *sb = pde_data(file_inode(seq->file));
3008 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
3011 return (void *) ((unsigned long) group);
3014 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
3016 struct super_block *sb = pde_data(file_inode(seq->file));
3017 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
3019 int err, buddy_loaded = 0;
3020 struct ext4_buddy e4b;
3021 struct ext4_group_info *grinfo;
3022 unsigned char blocksize_bits = min_t(unsigned char,
3023 sb->s_blocksize_bits,
3024 EXT4_MAX_BLOCK_LOG_SIZE);
3026 struct ext4_group_info info;
3027 ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2];
3032 seq_puts(seq, "#group: free frags first ["
3033 " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 "
3034 " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n");
3036 i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
3037 sizeof(struct ext4_group_info);
3039 grinfo = ext4_get_group_info(sb, group);
3042 /* Load the group info in memory only if not already loaded. */
3043 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
3044 err = ext4_mb_load_buddy(sb, group, &e4b);
3046 seq_printf(seq, "#%-5u: I/O error\n", group);
3052 memcpy(&sg, grinfo, i);
3055 ext4_mb_unload_buddy(&e4b);
3057 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
3058 sg.info.bb_fragments, sg.info.bb_first_free);
3059 for (i = 0; i <= 13; i++)
3060 seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ?
3061 sg.info.bb_counters[i] : 0);
3062 seq_puts(seq, " ]\n");
3067 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
3071 const struct seq_operations ext4_mb_seq_groups_ops = {
3072 .start = ext4_mb_seq_groups_start,
3073 .next = ext4_mb_seq_groups_next,
3074 .stop = ext4_mb_seq_groups_stop,
3075 .show = ext4_mb_seq_groups_show,
3078 int ext4_seq_mb_stats_show(struct seq_file *seq, void *offset)
3080 struct super_block *sb = seq->private;
3081 struct ext4_sb_info *sbi = EXT4_SB(sb);
3083 seq_puts(seq, "mballoc:\n");
3084 if (!sbi->s_mb_stats) {
3085 seq_puts(seq, "\tmb stats collection turned off.\n");
3088 "\tTo enable, please write \"1\" to sysfs file mb_stats.\n");
3091 seq_printf(seq, "\treqs: %u\n", atomic_read(&sbi->s_bal_reqs));
3092 seq_printf(seq, "\tsuccess: %u\n", atomic_read(&sbi->s_bal_success));
3094 seq_printf(seq, "\tgroups_scanned: %u\n",
3095 atomic_read(&sbi->s_bal_groups_scanned));
3097 /* CR_POWER2_ALIGNED stats */
3098 seq_puts(seq, "\tcr_p2_aligned_stats:\n");
3099 seq_printf(seq, "\t\thits: %llu\n",
3100 atomic64_read(&sbi->s_bal_cX_hits[CR_POWER2_ALIGNED]));
3102 seq, "\t\tgroups_considered: %llu\n",
3104 &sbi->s_bal_cX_groups_considered[CR_POWER2_ALIGNED]));
3105 seq_printf(seq, "\t\textents_scanned: %u\n",
3106 atomic_read(&sbi->s_bal_cX_ex_scanned[CR_POWER2_ALIGNED]));
3107 seq_printf(seq, "\t\tuseless_loops: %llu\n",
3108 atomic64_read(&sbi->s_bal_cX_failed[CR_POWER2_ALIGNED]));
3109 seq_printf(seq, "\t\tbad_suggestions: %u\n",
3110 atomic_read(&sbi->s_bal_p2_aligned_bad_suggestions));
3112 /* CR_GOAL_LEN_FAST stats */
3113 seq_puts(seq, "\tcr_goal_fast_stats:\n");
3114 seq_printf(seq, "\t\thits: %llu\n",
3115 atomic64_read(&sbi->s_bal_cX_hits[CR_GOAL_LEN_FAST]));
3116 seq_printf(seq, "\t\tgroups_considered: %llu\n",
3118 &sbi->s_bal_cX_groups_considered[CR_GOAL_LEN_FAST]));
3119 seq_printf(seq, "\t\textents_scanned: %u\n",
3120 atomic_read(&sbi->s_bal_cX_ex_scanned[CR_GOAL_LEN_FAST]));
3121 seq_printf(seq, "\t\tuseless_loops: %llu\n",
3122 atomic64_read(&sbi->s_bal_cX_failed[CR_GOAL_LEN_FAST]));
3123 seq_printf(seq, "\t\tbad_suggestions: %u\n",
3124 atomic_read(&sbi->s_bal_goal_fast_bad_suggestions));
3126 /* CR_BEST_AVAIL_LEN stats */
3127 seq_puts(seq, "\tcr_best_avail_stats:\n");
3128 seq_printf(seq, "\t\thits: %llu\n",
3129 atomic64_read(&sbi->s_bal_cX_hits[CR_BEST_AVAIL_LEN]));
3131 seq, "\t\tgroups_considered: %llu\n",
3133 &sbi->s_bal_cX_groups_considered[CR_BEST_AVAIL_LEN]));
3134 seq_printf(seq, "\t\textents_scanned: %u\n",
3135 atomic_read(&sbi->s_bal_cX_ex_scanned[CR_BEST_AVAIL_LEN]));
3136 seq_printf(seq, "\t\tuseless_loops: %llu\n",
3137 atomic64_read(&sbi->s_bal_cX_failed[CR_BEST_AVAIL_LEN]));
3138 seq_printf(seq, "\t\tbad_suggestions: %u\n",
3139 atomic_read(&sbi->s_bal_best_avail_bad_suggestions));
3141 /* CR_GOAL_LEN_SLOW stats */
3142 seq_puts(seq, "\tcr_goal_slow_stats:\n");
3143 seq_printf(seq, "\t\thits: %llu\n",
3144 atomic64_read(&sbi->s_bal_cX_hits[CR_GOAL_LEN_SLOW]));
3145 seq_printf(seq, "\t\tgroups_considered: %llu\n",
3147 &sbi->s_bal_cX_groups_considered[CR_GOAL_LEN_SLOW]));
3148 seq_printf(seq, "\t\textents_scanned: %u\n",
3149 atomic_read(&sbi->s_bal_cX_ex_scanned[CR_GOAL_LEN_SLOW]));
3150 seq_printf(seq, "\t\tuseless_loops: %llu\n",
3151 atomic64_read(&sbi->s_bal_cX_failed[CR_GOAL_LEN_SLOW]));
3153 /* CR_ANY_FREE stats */
3154 seq_puts(seq, "\tcr_any_free_stats:\n");
3155 seq_printf(seq, "\t\thits: %llu\n",
3156 atomic64_read(&sbi->s_bal_cX_hits[CR_ANY_FREE]));
3158 seq, "\t\tgroups_considered: %llu\n",
3159 atomic64_read(&sbi->s_bal_cX_groups_considered[CR_ANY_FREE]));
3160 seq_printf(seq, "\t\textents_scanned: %u\n",
3161 atomic_read(&sbi->s_bal_cX_ex_scanned[CR_ANY_FREE]));
3162 seq_printf(seq, "\t\tuseless_loops: %llu\n",
3163 atomic64_read(&sbi->s_bal_cX_failed[CR_ANY_FREE]));
3166 seq_printf(seq, "\textents_scanned: %u\n",
3167 atomic_read(&sbi->s_bal_ex_scanned));
3168 seq_printf(seq, "\t\tgoal_hits: %u\n", atomic_read(&sbi->s_bal_goals));
3169 seq_printf(seq, "\t\tlen_goal_hits: %u\n",
3170 atomic_read(&sbi->s_bal_len_goals));
3171 seq_printf(seq, "\t\t2^n_hits: %u\n", atomic_read(&sbi->s_bal_2orders));
3172 seq_printf(seq, "\t\tbreaks: %u\n", atomic_read(&sbi->s_bal_breaks));
3173 seq_printf(seq, "\t\tlost: %u\n", atomic_read(&sbi->s_mb_lost_chunks));
3174 seq_printf(seq, "\tbuddies_generated: %u/%u\n",
3175 atomic_read(&sbi->s_mb_buddies_generated),
3176 ext4_get_groups_count(sb));
3177 seq_printf(seq, "\tbuddies_time_used: %llu\n",
3178 atomic64_read(&sbi->s_mb_generation_time));
3179 seq_printf(seq, "\tpreallocated: %u\n",
3180 atomic_read(&sbi->s_mb_preallocated));
3181 seq_printf(seq, "\tdiscarded: %u\n", atomic_read(&sbi->s_mb_discarded));
3185 static void *ext4_mb_seq_structs_summary_start(struct seq_file *seq, loff_t *pos)
3186 __acquires(&EXT4_SB(sb)->s_mb_rb_lock)
3188 struct super_block *sb = pde_data(file_inode(seq->file));
3189 unsigned long position;
3191 if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb))
3193 position = *pos + 1;
3194 return (void *) ((unsigned long) position);
3197 static void *ext4_mb_seq_structs_summary_next(struct seq_file *seq, void *v, loff_t *pos)
3199 struct super_block *sb = pde_data(file_inode(seq->file));
3200 unsigned long position;
3203 if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb))
3205 position = *pos + 1;
3206 return (void *) ((unsigned long) position);
3209 static int ext4_mb_seq_structs_summary_show(struct seq_file *seq, void *v)
3211 struct super_block *sb = pde_data(file_inode(seq->file));
3212 struct ext4_sb_info *sbi = EXT4_SB(sb);
3213 unsigned long position = ((unsigned long) v);
3214 struct ext4_group_info *grp;
3218 if (position >= MB_NUM_ORDERS(sb)) {
3219 position -= MB_NUM_ORDERS(sb);
3221 seq_puts(seq, "avg_fragment_size_lists:\n");
3224 read_lock(&sbi->s_mb_avg_fragment_size_locks[position]);
3225 list_for_each_entry(grp, &sbi->s_mb_avg_fragment_size[position],
3226 bb_avg_fragment_size_node)
3228 read_unlock(&sbi->s_mb_avg_fragment_size_locks[position]);
3229 seq_printf(seq, "\tlist_order_%u_groups: %u\n",
3230 (unsigned int)position, count);
3234 if (position == 0) {
3235 seq_printf(seq, "optimize_scan: %d\n",
3236 test_opt2(sb, MB_OPTIMIZE_SCAN) ? 1 : 0);
3237 seq_puts(seq, "max_free_order_lists:\n");
3240 read_lock(&sbi->s_mb_largest_free_orders_locks[position]);
3241 list_for_each_entry(grp, &sbi->s_mb_largest_free_orders[position],
3242 bb_largest_free_order_node)
3244 read_unlock(&sbi->s_mb_largest_free_orders_locks[position]);
3245 seq_printf(seq, "\tlist_order_%u_groups: %u\n",
3246 (unsigned int)position, count);
3251 static void ext4_mb_seq_structs_summary_stop(struct seq_file *seq, void *v)
3255 const struct seq_operations ext4_mb_seq_structs_summary_ops = {
3256 .start = ext4_mb_seq_structs_summary_start,
3257 .next = ext4_mb_seq_structs_summary_next,
3258 .stop = ext4_mb_seq_structs_summary_stop,
3259 .show = ext4_mb_seq_structs_summary_show,
3262 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
3264 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
3265 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
3272 * Allocate the top-level s_group_info array for the specified number
3275 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
3277 struct ext4_sb_info *sbi = EXT4_SB(sb);
3279 struct ext4_group_info ***old_groupinfo, ***new_groupinfo;
3281 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
3282 EXT4_DESC_PER_BLOCK_BITS(sb);
3283 if (size <= sbi->s_group_info_size)
3286 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
3287 new_groupinfo = kvzalloc(size, GFP_KERNEL);
3288 if (!new_groupinfo) {
3289 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
3293 old_groupinfo = rcu_dereference(sbi->s_group_info);
3295 memcpy(new_groupinfo, old_groupinfo,
3296 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
3298 rcu_assign_pointer(sbi->s_group_info, new_groupinfo);
3299 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
3301 ext4_kvfree_array_rcu(old_groupinfo);
3302 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
3303 sbi->s_group_info_size);
3307 /* Create and initialize ext4_group_info data for the given group. */
3308 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
3309 struct ext4_group_desc *desc)
3313 int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb);
3314 struct ext4_sb_info *sbi = EXT4_SB(sb);
3315 struct ext4_group_info **meta_group_info;
3316 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3319 * First check if this group is the first of a reserved block.
3320 * If it's true, we have to allocate a new table of pointers
3321 * to ext4_group_info structures
3323 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
3324 metalen = sizeof(*meta_group_info) <<
3325 EXT4_DESC_PER_BLOCK_BITS(sb);
3326 meta_group_info = kmalloc(metalen, GFP_NOFS);
3327 if (meta_group_info == NULL) {
3328 ext4_msg(sb, KERN_ERR, "can't allocate mem "
3329 "for a buddy group");
3333 rcu_dereference(sbi->s_group_info)[idx] = meta_group_info;
3337 meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx);
3338 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
3340 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
3341 if (meta_group_info[i] == NULL) {
3342 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
3343 goto exit_group_info;
3345 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
3346 &(meta_group_info[i]->bb_state));
3349 * initialize bb_free to be able to skip
3350 * empty groups without initialization
3352 if (ext4_has_group_desc_csum(sb) &&
3353 (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
3354 meta_group_info[i]->bb_free =
3355 ext4_free_clusters_after_init(sb, group, desc);
3357 meta_group_info[i]->bb_free =
3358 ext4_free_group_clusters(sb, desc);
3361 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
3362 init_rwsem(&meta_group_info[i]->alloc_sem);
3363 meta_group_info[i]->bb_free_root = RB_ROOT;
3364 INIT_LIST_HEAD(&meta_group_info[i]->bb_largest_free_order_node);
3365 INIT_LIST_HEAD(&meta_group_info[i]->bb_avg_fragment_size_node);
3366 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
3367 meta_group_info[i]->bb_avg_fragment_size_order = -1; /* uninit */
3368 meta_group_info[i]->bb_group = group;
3370 mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group);
3374 /* If a meta_group_info table has been allocated, release it now */
3375 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
3376 struct ext4_group_info ***group_info;
3379 group_info = rcu_dereference(sbi->s_group_info);
3380 kfree(group_info[idx]);
3381 group_info[idx] = NULL;
3385 } /* ext4_mb_add_groupinfo */
3387 static int ext4_mb_init_backend(struct super_block *sb)
3389 ext4_group_t ngroups = ext4_get_groups_count(sb);
3391 struct ext4_sb_info *sbi = EXT4_SB(sb);
3393 struct ext4_group_desc *desc;
3394 struct ext4_group_info ***group_info;
3395 struct kmem_cache *cachep;
3397 err = ext4_mb_alloc_groupinfo(sb, ngroups);
3401 sbi->s_buddy_cache = new_inode(sb);
3402 if (sbi->s_buddy_cache == NULL) {
3403 ext4_msg(sb, KERN_ERR, "can't get new inode");
3406 /* To avoid potentially colliding with an valid on-disk inode number,
3407 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
3408 * not in the inode hash, so it should never be found by iget(), but
3409 * this will avoid confusion if it ever shows up during debugging. */
3410 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
3411 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
3412 for (i = 0; i < ngroups; i++) {
3414 desc = ext4_get_group_desc(sb, i, NULL);
3416 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
3419 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
3423 if (ext4_has_feature_flex_bg(sb)) {
3424 /* a single flex group is supposed to be read by a single IO.
3425 * 2 ^ s_log_groups_per_flex != UINT_MAX as s_mb_prefetch is
3426 * unsigned integer, so the maximum shift is 32.
3428 if (sbi->s_es->s_log_groups_per_flex >= 32) {
3429 ext4_msg(sb, KERN_ERR, "too many log groups per flexible block group");
3432 sbi->s_mb_prefetch = min_t(uint, 1 << sbi->s_es->s_log_groups_per_flex,
3433 BLK_MAX_SEGMENT_SIZE >> (sb->s_blocksize_bits - 9));
3434 sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */
3436 sbi->s_mb_prefetch = 32;
3438 if (sbi->s_mb_prefetch > ext4_get_groups_count(sb))
3439 sbi->s_mb_prefetch = ext4_get_groups_count(sb);
3440 /* now many real IOs to prefetch within a single allocation at cr=0
3441 * given cr=0 is an CPU-related optimization we shouldn't try to
3442 * load too many groups, at some point we should start to use what
3443 * we've got in memory.
3444 * with an average random access time 5ms, it'd take a second to get
3445 * 200 groups (* N with flex_bg), so let's make this limit 4
3447 sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4;
3448 if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb))
3449 sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb);
3454 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3456 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
3459 kmem_cache_free(cachep, grp);
3461 i = sbi->s_group_info_size;
3463 group_info = rcu_dereference(sbi->s_group_info);
3465 kfree(group_info[i]);
3467 iput(sbi->s_buddy_cache);
3470 kvfree(rcu_dereference(sbi->s_group_info));
3475 static void ext4_groupinfo_destroy_slabs(void)
3479 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
3480 kmem_cache_destroy(ext4_groupinfo_caches[i]);
3481 ext4_groupinfo_caches[i] = NULL;
3485 static int ext4_groupinfo_create_slab(size_t size)
3487 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
3489 int blocksize_bits = order_base_2(size);
3490 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
3491 struct kmem_cache *cachep;
3493 if (cache_index >= NR_GRPINFO_CACHES)
3496 if (unlikely(cache_index < 0))
3499 mutex_lock(&ext4_grpinfo_slab_create_mutex);
3500 if (ext4_groupinfo_caches[cache_index]) {
3501 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
3502 return 0; /* Already created */
3505 slab_size = offsetof(struct ext4_group_info,
3506 bb_counters[blocksize_bits + 2]);
3508 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
3509 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
3512 ext4_groupinfo_caches[cache_index] = cachep;
3514 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
3517 "EXT4-fs: no memory for groupinfo slab cache\n");
3524 static void ext4_discard_work(struct work_struct *work)
3526 struct ext4_sb_info *sbi = container_of(work,
3527 struct ext4_sb_info, s_discard_work);
3528 struct super_block *sb = sbi->s_sb;
3529 struct ext4_free_data *fd, *nfd;
3530 struct ext4_buddy e4b;
3531 LIST_HEAD(discard_list);
3532 ext4_group_t grp, load_grp;
3535 spin_lock(&sbi->s_md_lock);
3536 list_splice_init(&sbi->s_discard_list, &discard_list);
3537 spin_unlock(&sbi->s_md_lock);
3539 load_grp = UINT_MAX;
3540 list_for_each_entry_safe(fd, nfd, &discard_list, efd_list) {
3542 * If filesystem is umounting or no memory or suffering
3543 * from no space, give up the discard
3545 if ((sb->s_flags & SB_ACTIVE) && !err &&
3546 !atomic_read(&sbi->s_retry_alloc_pending)) {
3547 grp = fd->efd_group;
3548 if (grp != load_grp) {
3549 if (load_grp != UINT_MAX)
3550 ext4_mb_unload_buddy(&e4b);
3552 err = ext4_mb_load_buddy(sb, grp, &e4b);
3554 kmem_cache_free(ext4_free_data_cachep, fd);
3555 load_grp = UINT_MAX;
3562 ext4_lock_group(sb, grp);
3563 ext4_try_to_trim_range(sb, &e4b, fd->efd_start_cluster,
3564 fd->efd_start_cluster + fd->efd_count - 1, 1);
3565 ext4_unlock_group(sb, grp);
3567 kmem_cache_free(ext4_free_data_cachep, fd);
3570 if (load_grp != UINT_MAX)
3571 ext4_mb_unload_buddy(&e4b);
3574 int ext4_mb_init(struct super_block *sb)
3576 struct ext4_sb_info *sbi = EXT4_SB(sb);
3578 unsigned offset, offset_incr;
3582 i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_offsets);
3584 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
3585 if (sbi->s_mb_offsets == NULL) {
3590 i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_maxs);
3591 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
3592 if (sbi->s_mb_maxs == NULL) {
3597 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
3601 /* order 0 is regular bitmap */
3602 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
3603 sbi->s_mb_offsets[0] = 0;
3607 offset_incr = 1 << (sb->s_blocksize_bits - 1);
3608 max = sb->s_blocksize << 2;
3610 sbi->s_mb_offsets[i] = offset;
3611 sbi->s_mb_maxs[i] = max;
3612 offset += offset_incr;
3613 offset_incr = offset_incr >> 1;
3616 } while (i < MB_NUM_ORDERS(sb));
3618 sbi->s_mb_avg_fragment_size =
3619 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head),
3621 if (!sbi->s_mb_avg_fragment_size) {
3625 sbi->s_mb_avg_fragment_size_locks =
3626 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t),
3628 if (!sbi->s_mb_avg_fragment_size_locks) {
3632 for (i = 0; i < MB_NUM_ORDERS(sb); i++) {
3633 INIT_LIST_HEAD(&sbi->s_mb_avg_fragment_size[i]);
3634 rwlock_init(&sbi->s_mb_avg_fragment_size_locks[i]);
3636 sbi->s_mb_largest_free_orders =
3637 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head),
3639 if (!sbi->s_mb_largest_free_orders) {
3643 sbi->s_mb_largest_free_orders_locks =
3644 kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t),
3646 if (!sbi->s_mb_largest_free_orders_locks) {
3650 for (i = 0; i < MB_NUM_ORDERS(sb); i++) {
3651 INIT_LIST_HEAD(&sbi->s_mb_largest_free_orders[i]);
3652 rwlock_init(&sbi->s_mb_largest_free_orders_locks[i]);
3655 spin_lock_init(&sbi->s_md_lock);
3656 sbi->s_mb_free_pending = 0;
3657 INIT_LIST_HEAD(&sbi->s_freed_data_list[0]);
3658 INIT_LIST_HEAD(&sbi->s_freed_data_list[1]);
3659 INIT_LIST_HEAD(&sbi->s_discard_list);
3660 INIT_WORK(&sbi->s_discard_work, ext4_discard_work);
3661 atomic_set(&sbi->s_retry_alloc_pending, 0);
3663 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
3664 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
3665 sbi->s_mb_stats = MB_DEFAULT_STATS;
3666 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
3667 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
3668 sbi->s_mb_best_avail_max_trim_order = MB_DEFAULT_BEST_AVAIL_TRIM_ORDER;
3671 * The default group preallocation is 512, which for 4k block
3672 * sizes translates to 2 megabytes. However for bigalloc file
3673 * systems, this is probably too big (i.e, if the cluster size
3674 * is 1 megabyte, then group preallocation size becomes half a
3675 * gigabyte!). As a default, we will keep a two megabyte
3676 * group pralloc size for cluster sizes up to 64k, and after
3677 * that, we will force a minimum group preallocation size of
3678 * 32 clusters. This translates to 8 megs when the cluster
3679 * size is 256k, and 32 megs when the cluster size is 1 meg,
3680 * which seems reasonable as a default.
3682 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
3683 sbi->s_cluster_bits, 32);
3685 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
3686 * to the lowest multiple of s_stripe which is bigger than
3687 * the s_mb_group_prealloc as determined above. We want
3688 * the preallocation size to be an exact multiple of the
3689 * RAID stripe size so that preallocations don't fragment
3692 if (sbi->s_stripe > 1) {
3693 sbi->s_mb_group_prealloc = roundup(
3694 sbi->s_mb_group_prealloc, EXT4_B2C(sbi, sbi->s_stripe));
3697 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
3698 if (sbi->s_locality_groups == NULL) {
3702 for_each_possible_cpu(i) {
3703 struct ext4_locality_group *lg;
3704 lg = per_cpu_ptr(sbi->s_locality_groups, i);
3705 mutex_init(&lg->lg_mutex);
3706 for (j = 0; j < PREALLOC_TB_SIZE; j++)
3707 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
3708 spin_lock_init(&lg->lg_prealloc_lock);
3711 if (bdev_nonrot(sb->s_bdev))
3712 sbi->s_mb_max_linear_groups = 0;
3714 sbi->s_mb_max_linear_groups = MB_DEFAULT_LINEAR_LIMIT;
3715 /* init file for buddy data */
3716 ret = ext4_mb_init_backend(sb);
3718 goto out_free_locality_groups;
3722 out_free_locality_groups:
3723 free_percpu(sbi->s_locality_groups);
3724 sbi->s_locality_groups = NULL;
3726 kfree(sbi->s_mb_avg_fragment_size);
3727 kfree(sbi->s_mb_avg_fragment_size_locks);
3728 kfree(sbi->s_mb_largest_free_orders);
3729 kfree(sbi->s_mb_largest_free_orders_locks);
3730 kfree(sbi->s_mb_offsets);
3731 sbi->s_mb_offsets = NULL;
3732 kfree(sbi->s_mb_maxs);
3733 sbi->s_mb_maxs = NULL;
3737 /* need to called with the ext4 group lock held */
3738 static int ext4_mb_cleanup_pa(struct ext4_group_info *grp)
3740 struct ext4_prealloc_space *pa;
3741 struct list_head *cur, *tmp;
3744 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
3745 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3746 list_del(&pa->pa_group_list);
3748 kmem_cache_free(ext4_pspace_cachep, pa);
3753 void ext4_mb_release(struct super_block *sb)
3755 ext4_group_t ngroups = ext4_get_groups_count(sb);
3757 int num_meta_group_infos;
3758 struct ext4_group_info *grinfo, ***group_info;
3759 struct ext4_sb_info *sbi = EXT4_SB(sb);
3760 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
3763 if (test_opt(sb, DISCARD)) {
3765 * wait the discard work to drain all of ext4_free_data
3767 flush_work(&sbi->s_discard_work);
3768 WARN_ON_ONCE(!list_empty(&sbi->s_discard_list));
3771 if (sbi->s_group_info) {
3772 for (i = 0; i < ngroups; i++) {
3774 grinfo = ext4_get_group_info(sb, i);
3777 mb_group_bb_bitmap_free(grinfo);
3778 ext4_lock_group(sb, i);
3779 count = ext4_mb_cleanup_pa(grinfo);
3781 mb_debug(sb, "mballoc: %d PAs left\n",
3783 ext4_unlock_group(sb, i);
3784 kmem_cache_free(cachep, grinfo);
3786 num_meta_group_infos = (ngroups +
3787 EXT4_DESC_PER_BLOCK(sb) - 1) >>
3788 EXT4_DESC_PER_BLOCK_BITS(sb);
3790 group_info = rcu_dereference(sbi->s_group_info);
3791 for (i = 0; i < num_meta_group_infos; i++)
3792 kfree(group_info[i]);
3796 kfree(sbi->s_mb_avg_fragment_size);
3797 kfree(sbi->s_mb_avg_fragment_size_locks);
3798 kfree(sbi->s_mb_largest_free_orders);
3799 kfree(sbi->s_mb_largest_free_orders_locks);
3800 kfree(sbi->s_mb_offsets);
3801 kfree(sbi->s_mb_maxs);
3802 iput(sbi->s_buddy_cache);
3803 if (sbi->s_mb_stats) {
3804 ext4_msg(sb, KERN_INFO,
3805 "mballoc: %u blocks %u reqs (%u success)",
3806 atomic_read(&sbi->s_bal_allocated),
3807 atomic_read(&sbi->s_bal_reqs),
3808 atomic_read(&sbi->s_bal_success));
3809 ext4_msg(sb, KERN_INFO,
3810 "mballoc: %u extents scanned, %u groups scanned, %u goal hits, "
3811 "%u 2^N hits, %u breaks, %u lost",
3812 atomic_read(&sbi->s_bal_ex_scanned),
3813 atomic_read(&sbi->s_bal_groups_scanned),
3814 atomic_read(&sbi->s_bal_goals),
3815 atomic_read(&sbi->s_bal_2orders),
3816 atomic_read(&sbi->s_bal_breaks),
3817 atomic_read(&sbi->s_mb_lost_chunks));
3818 ext4_msg(sb, KERN_INFO,
3819 "mballoc: %u generated and it took %llu",
3820 atomic_read(&sbi->s_mb_buddies_generated),
3821 atomic64_read(&sbi->s_mb_generation_time));
3822 ext4_msg(sb, KERN_INFO,
3823 "mballoc: %u preallocated, %u discarded",
3824 atomic_read(&sbi->s_mb_preallocated),
3825 atomic_read(&sbi->s_mb_discarded));
3828 free_percpu(sbi->s_locality_groups);
3831 static inline int ext4_issue_discard(struct super_block *sb,
3832 ext4_group_t block_group, ext4_grpblk_t cluster, int count,
3835 ext4_fsblk_t discard_block;
3837 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
3838 ext4_group_first_block_no(sb, block_group));
3839 count = EXT4_C2B(EXT4_SB(sb), count);
3840 trace_ext4_discard_blocks(sb,
3841 (unsigned long long) discard_block, count);
3843 return __blkdev_issue_discard(sb->s_bdev,
3844 (sector_t)discard_block << (sb->s_blocksize_bits - 9),
3845 (sector_t)count << (sb->s_blocksize_bits - 9),
3848 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
3851 static void ext4_free_data_in_buddy(struct super_block *sb,
3852 struct ext4_free_data *entry)
3854 struct ext4_buddy e4b;
3855 struct ext4_group_info *db;
3858 mb_debug(sb, "gonna free %u blocks in group %u (0x%p):",
3859 entry->efd_count, entry->efd_group, entry);
3861 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
3862 /* we expect to find existing buddy because it's pinned */
3865 spin_lock(&EXT4_SB(sb)->s_md_lock);
3866 EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count;
3867 spin_unlock(&EXT4_SB(sb)->s_md_lock);
3870 /* there are blocks to put in buddy to make them really free */
3871 count += entry->efd_count;
3872 ext4_lock_group(sb, entry->efd_group);
3873 /* Take it out of per group rb tree */
3874 rb_erase(&entry->efd_node, &(db->bb_free_root));
3875 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
3878 * Clear the trimmed flag for the group so that the next
3879 * ext4_trim_fs can trim it.
3880 * If the volume is mounted with -o discard, online discard
3881 * is supported and the free blocks will be trimmed online.
3883 if (!test_opt(sb, DISCARD))
3884 EXT4_MB_GRP_CLEAR_TRIMMED(db);
3886 if (!db->bb_free_root.rb_node) {
3887 /* No more items in the per group rb tree
3888 * balance refcounts from ext4_mb_free_metadata()
3890 put_page(e4b.bd_buddy_page);
3891 put_page(e4b.bd_bitmap_page);
3893 ext4_unlock_group(sb, entry->efd_group);
3894 ext4_mb_unload_buddy(&e4b);
3896 mb_debug(sb, "freed %d blocks in 1 structures\n", count);
3900 * This function is called by the jbd2 layer once the commit has finished,
3901 * so we know we can free the blocks that were released with that commit.
3903 void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid)
3905 struct ext4_sb_info *sbi = EXT4_SB(sb);
3906 struct ext4_free_data *entry, *tmp;
3907 LIST_HEAD(freed_data_list);
3908 struct list_head *s_freed_head = &sbi->s_freed_data_list[commit_tid & 1];
3911 list_replace_init(s_freed_head, &freed_data_list);
3913 list_for_each_entry(entry, &freed_data_list, efd_list)
3914 ext4_free_data_in_buddy(sb, entry);
3916 if (test_opt(sb, DISCARD)) {
3917 spin_lock(&sbi->s_md_lock);
3918 wake = list_empty(&sbi->s_discard_list);
3919 list_splice_tail(&freed_data_list, &sbi->s_discard_list);
3920 spin_unlock(&sbi->s_md_lock);
3922 queue_work(system_unbound_wq, &sbi->s_discard_work);
3924 list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list)
3925 kmem_cache_free(ext4_free_data_cachep, entry);
3929 int __init ext4_init_mballoc(void)
3931 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
3932 SLAB_RECLAIM_ACCOUNT);
3933 if (ext4_pspace_cachep == NULL)
3936 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
3937 SLAB_RECLAIM_ACCOUNT);
3938 if (ext4_ac_cachep == NULL)
3941 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
3942 SLAB_RECLAIM_ACCOUNT);
3943 if (ext4_free_data_cachep == NULL)
3949 kmem_cache_destroy(ext4_ac_cachep);
3951 kmem_cache_destroy(ext4_pspace_cachep);
3956 void ext4_exit_mballoc(void)
3959 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
3960 * before destroying the slab cache.
3963 kmem_cache_destroy(ext4_pspace_cachep);
3964 kmem_cache_destroy(ext4_ac_cachep);
3965 kmem_cache_destroy(ext4_free_data_cachep);
3966 ext4_groupinfo_destroy_slabs();
3969 #define EXT4_MB_BITMAP_MARKED_CHECK 0x0001
3970 #define EXT4_MB_SYNC_UPDATE 0x0002
3972 ext4_mb_mark_context(handle_t *handle, struct super_block *sb, bool state,
3973 ext4_group_t group, ext4_grpblk_t blkoff,
3974 ext4_grpblk_t len, int flags, ext4_grpblk_t *ret_changed)
3976 struct ext4_sb_info *sbi = EXT4_SB(sb);
3977 struct buffer_head *bitmap_bh = NULL;
3978 struct ext4_group_desc *gdp;
3979 struct buffer_head *gdp_bh;
3981 unsigned int i, already, changed = len;
3983 KUNIT_STATIC_STUB_REDIRECT(ext4_mb_mark_context,
3984 handle, sb, state, group, blkoff, len,
3985 flags, ret_changed);
3989 bitmap_bh = ext4_read_block_bitmap(sb, group);
3990 if (IS_ERR(bitmap_bh))
3991 return PTR_ERR(bitmap_bh);
3994 BUFFER_TRACE(bitmap_bh, "getting write access");
3995 err = ext4_journal_get_write_access(handle, sb, bitmap_bh,
4002 gdp = ext4_get_group_desc(sb, group, &gdp_bh);
4007 BUFFER_TRACE(gdp_bh, "get_write_access");
4008 err = ext4_journal_get_write_access(handle, sb, gdp_bh,
4014 ext4_lock_group(sb, group);
4015 if (ext4_has_group_desc_csum(sb) &&
4016 (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
4017 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
4018 ext4_free_group_clusters_set(sb, gdp,
4019 ext4_free_clusters_after_init(sb, group, gdp));
4022 if (flags & EXT4_MB_BITMAP_MARKED_CHECK) {
4024 for (i = 0; i < len; i++)
4025 if (mb_test_bit(blkoff + i, bitmap_bh->b_data) ==
4028 changed = len - already;
4032 mb_set_bits(bitmap_bh->b_data, blkoff, len);
4033 ext4_free_group_clusters_set(sb, gdp,
4034 ext4_free_group_clusters(sb, gdp) - changed);
4036 mb_clear_bits(bitmap_bh->b_data, blkoff, len);
4037 ext4_free_group_clusters_set(sb, gdp,
4038 ext4_free_group_clusters(sb, gdp) + changed);
4041 ext4_block_bitmap_csum_set(sb, gdp, bitmap_bh);
4042 ext4_group_desc_csum_set(sb, group, gdp);
4043 ext4_unlock_group(sb, group);
4045 *ret_changed = changed;
4047 if (sbi->s_log_groups_per_flex) {
4048 ext4_group_t flex_group = ext4_flex_group(sbi, group);
4049 struct flex_groups *fg = sbi_array_rcu_deref(sbi,
4050 s_flex_groups, flex_group);
4053 atomic64_sub(changed, &fg->free_clusters);
4055 atomic64_add(changed, &fg->free_clusters);
4058 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4061 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
4065 if (flags & EXT4_MB_SYNC_UPDATE) {
4066 sync_dirty_buffer(bitmap_bh);
4067 sync_dirty_buffer(gdp_bh);
4076 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
4077 * Returns 0 if success or error code
4079 static noinline_for_stack int
4080 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
4081 handle_t *handle, unsigned int reserv_clstrs)
4083 struct ext4_group_desc *gdp;
4084 struct ext4_sb_info *sbi;
4085 struct super_block *sb;
4089 ext4_grpblk_t changed;
4091 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
4092 BUG_ON(ac->ac_b_ex.fe_len <= 0);
4097 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, NULL);
4100 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
4101 ext4_free_group_clusters(sb, gdp));
4103 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4104 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4105 if (!ext4_inode_block_valid(ac->ac_inode, block, len)) {
4106 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
4107 "fs metadata", block, block+len);
4108 /* File system mounted not to panic on error
4109 * Fix the bitmap and return EFSCORRUPTED
4110 * We leak some of the blocks here.
4112 err = ext4_mb_mark_context(handle, sb, true,
4113 ac->ac_b_ex.fe_group,
4114 ac->ac_b_ex.fe_start,
4118 err = -EFSCORRUPTED;
4122 #ifdef AGGRESSIVE_CHECK
4123 flags |= EXT4_MB_BITMAP_MARKED_CHECK;
4125 err = ext4_mb_mark_context(handle, sb, true, ac->ac_b_ex.fe_group,
4126 ac->ac_b_ex.fe_start, ac->ac_b_ex.fe_len,
4129 if (err && changed == 0)
4132 #ifdef AGGRESSIVE_CHECK
4133 BUG_ON(changed != ac->ac_b_ex.fe_len);
4135 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
4137 * Now reduce the dirty block count also. Should not go negative
4139 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
4140 /* release all the reserved blocks if non delalloc */
4141 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4148 * Idempotent helper for Ext4 fast commit replay path to set the state of
4149 * blocks in bitmaps and update counters.
4151 void ext4_mb_mark_bb(struct super_block *sb, ext4_fsblk_t block,
4152 int len, bool state)
4154 struct ext4_sb_info *sbi = EXT4_SB(sb);
4156 ext4_grpblk_t blkoff;
4158 unsigned int clen, thisgrp_len;
4161 ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
4164 * Check to see if we are freeing blocks across a group
4166 * In case of flex_bg, this can happen that (block, len) may
4167 * span across more than one group. In that case we need to
4168 * get the corresponding group metadata to work with.
4169 * For this we have goto again loop.
4171 thisgrp_len = min_t(unsigned int, (unsigned int)len,
4172 EXT4_BLOCKS_PER_GROUP(sb) - EXT4_C2B(sbi, blkoff));
4173 clen = EXT4_NUM_B2C(sbi, thisgrp_len);
4175 if (!ext4_sb_block_valid(sb, NULL, block, thisgrp_len)) {
4176 ext4_error(sb, "Marking blocks in system zone - "
4177 "Block = %llu, len = %u",
4178 block, thisgrp_len);
4182 err = ext4_mb_mark_context(NULL, sb, state,
4183 group, blkoff, clen,
4184 EXT4_MB_BITMAP_MARKED_CHECK |
4185 EXT4_MB_SYNC_UPDATE,
4190 block += thisgrp_len;
4197 * here we normalize request for locality group
4198 * Group request are normalized to s_mb_group_prealloc, which goes to
4199 * s_strip if we set the same via mount option.
4200 * s_mb_group_prealloc can be configured via
4201 * /sys/fs/ext4/<partition>/mb_group_prealloc
4203 * XXX: should we try to preallocate more than the group has now?
4205 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
4207 struct super_block *sb = ac->ac_sb;
4208 struct ext4_locality_group *lg = ac->ac_lg;
4211 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
4212 mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len);
4216 * This function returns the next element to look at during inode
4217 * PA rbtree walk. We assume that we have held the inode PA rbtree lock
4218 * (ei->i_prealloc_lock)
4220 * new_start The start of the range we want to compare
4221 * cur_start The existing start that we are comparing against
4222 * node The node of the rb_tree
4224 static inline struct rb_node*
4225 ext4_mb_pa_rb_next_iter(ext4_lblk_t new_start, ext4_lblk_t cur_start, struct rb_node *node)
4227 if (new_start < cur_start)
4228 return node->rb_left;
4230 return node->rb_right;
4234 ext4_mb_pa_assert_overlap(struct ext4_allocation_context *ac,
4235 ext4_lblk_t start, loff_t end)
4237 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4238 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
4239 struct ext4_prealloc_space *tmp_pa;
4240 ext4_lblk_t tmp_pa_start;
4242 struct rb_node *iter;
4244 read_lock(&ei->i_prealloc_lock);
4245 for (iter = ei->i_prealloc_node.rb_node; iter;
4246 iter = ext4_mb_pa_rb_next_iter(start, tmp_pa_start, iter)) {
4247 tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
4248 pa_node.inode_node);
4249 tmp_pa_start = tmp_pa->pa_lstart;
4250 tmp_pa_end = pa_logical_end(sbi, tmp_pa);
4252 spin_lock(&tmp_pa->pa_lock);
4253 if (tmp_pa->pa_deleted == 0)
4254 BUG_ON(!(start >= tmp_pa_end || end <= tmp_pa_start));
4255 spin_unlock(&tmp_pa->pa_lock);
4257 read_unlock(&ei->i_prealloc_lock);
4261 * Given an allocation context "ac" and a range "start", "end", check
4262 * and adjust boundaries if the range overlaps with any of the existing
4263 * preallocatoins stored in the corresponding inode of the allocation context.
4266 * ac allocation context
4267 * start start of the new range
4268 * end end of the new range
4271 ext4_mb_pa_adjust_overlap(struct ext4_allocation_context *ac,
4272 ext4_lblk_t *start, loff_t *end)
4274 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
4275 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4276 struct ext4_prealloc_space *tmp_pa = NULL, *left_pa = NULL, *right_pa = NULL;
4277 struct rb_node *iter;
4278 ext4_lblk_t new_start, tmp_pa_start, right_pa_start = -1;
4279 loff_t new_end, tmp_pa_end, left_pa_end = -1;
4285 * Adjust the normalized range so that it doesn't overlap with any
4286 * existing preallocated blocks(PAs). Make sure to hold the rbtree lock
4287 * so it doesn't change underneath us.
4289 read_lock(&ei->i_prealloc_lock);
4291 /* Step 1: find any one immediate neighboring PA of the normalized range */
4292 for (iter = ei->i_prealloc_node.rb_node; iter;
4293 iter = ext4_mb_pa_rb_next_iter(ac->ac_o_ex.fe_logical,
4294 tmp_pa_start, iter)) {
4295 tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
4296 pa_node.inode_node);
4297 tmp_pa_start = tmp_pa->pa_lstart;
4298 tmp_pa_end = pa_logical_end(sbi, tmp_pa);
4300 /* PA must not overlap original request */
4301 spin_lock(&tmp_pa->pa_lock);
4302 if (tmp_pa->pa_deleted == 0)
4303 BUG_ON(!(ac->ac_o_ex.fe_logical >= tmp_pa_end ||
4304 ac->ac_o_ex.fe_logical < tmp_pa_start));
4305 spin_unlock(&tmp_pa->pa_lock);
4309 * Step 2: check if the found PA is left or right neighbor and
4310 * get the other neighbor
4313 if (tmp_pa->pa_lstart < ac->ac_o_ex.fe_logical) {
4314 struct rb_node *tmp;
4317 tmp = rb_next(&left_pa->pa_node.inode_node);
4319 right_pa = rb_entry(tmp,
4320 struct ext4_prealloc_space,
4321 pa_node.inode_node);
4324 struct rb_node *tmp;
4327 tmp = rb_prev(&right_pa->pa_node.inode_node);
4329 left_pa = rb_entry(tmp,
4330 struct ext4_prealloc_space,
4331 pa_node.inode_node);
4336 /* Step 3: get the non deleted neighbors */
4338 for (iter = &left_pa->pa_node.inode_node;;
4339 iter = rb_prev(iter)) {
4345 tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
4346 pa_node.inode_node);
4348 spin_lock(&tmp_pa->pa_lock);
4349 if (tmp_pa->pa_deleted == 0) {
4350 spin_unlock(&tmp_pa->pa_lock);
4353 spin_unlock(&tmp_pa->pa_lock);
4358 for (iter = &right_pa->pa_node.inode_node;;
4359 iter = rb_next(iter)) {
4365 tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
4366 pa_node.inode_node);
4368 spin_lock(&tmp_pa->pa_lock);
4369 if (tmp_pa->pa_deleted == 0) {
4370 spin_unlock(&tmp_pa->pa_lock);
4373 spin_unlock(&tmp_pa->pa_lock);
4378 left_pa_end = pa_logical_end(sbi, left_pa);
4379 BUG_ON(left_pa_end > ac->ac_o_ex.fe_logical);
4383 right_pa_start = right_pa->pa_lstart;
4384 BUG_ON(right_pa_start <= ac->ac_o_ex.fe_logical);
4387 /* Step 4: trim our normalized range to not overlap with the neighbors */
4389 if (left_pa_end > new_start)
4390 new_start = left_pa_end;
4394 if (right_pa_start < new_end)
4395 new_end = right_pa_start;
4397 read_unlock(&ei->i_prealloc_lock);
4399 /* XXX: extra loop to check we really don't overlap preallocations */
4400 ext4_mb_pa_assert_overlap(ac, new_start, new_end);
4407 * Normalization means making request better in terms of
4408 * size and alignment
4410 static noinline_for_stack void
4411 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
4412 struct ext4_allocation_request *ar)
4414 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4415 struct ext4_super_block *es = sbi->s_es;
4417 loff_t size, start_off, end;
4418 loff_t orig_size __maybe_unused;
4421 /* do normalize only data requests, metadata requests
4422 do not need preallocation */
4423 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4426 /* sometime caller may want exact blocks */
4427 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4430 /* caller may indicate that preallocation isn't
4431 * required (it's a tail, for example) */
4432 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
4435 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
4436 ext4_mb_normalize_group_request(ac);
4440 bsbits = ac->ac_sb->s_blocksize_bits;
4442 /* first, let's learn actual file size
4443 * given current request is allocated */
4444 size = extent_logical_end(sbi, &ac->ac_o_ex);
4445 size = size << bsbits;
4446 if (size < i_size_read(ac->ac_inode))
4447 size = i_size_read(ac->ac_inode);
4450 /* max size of free chunks */
4453 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
4454 (req <= (size) || max <= (chunk_size))
4456 /* first, try to predict filesize */
4457 /* XXX: should this table be tunable? */
4459 if (size <= 16 * 1024) {
4461 } else if (size <= 32 * 1024) {
4463 } else if (size <= 64 * 1024) {
4465 } else if (size <= 128 * 1024) {
4467 } else if (size <= 256 * 1024) {
4469 } else if (size <= 512 * 1024) {
4471 } else if (size <= 1024 * 1024) {
4473 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
4474 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4475 (21 - bsbits)) << 21;
4476 size = 2 * 1024 * 1024;
4477 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
4478 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4479 (22 - bsbits)) << 22;
4480 size = 4 * 1024 * 1024;
4481 } else if (NRL_CHECK_SIZE(EXT4_C2B(sbi, ac->ac_o_ex.fe_len),
4482 (8<<20)>>bsbits, max, 8 * 1024)) {
4483 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
4484 (23 - bsbits)) << 23;
4485 size = 8 * 1024 * 1024;
4487 start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
4488 size = (loff_t) EXT4_C2B(sbi,
4489 ac->ac_o_ex.fe_len) << bsbits;
4491 size = size >> bsbits;
4492 start = start_off >> bsbits;
4495 * For tiny groups (smaller than 8MB) the chosen allocation
4496 * alignment may be larger than group size. Make sure the
4497 * alignment does not move allocation to a different group which
4498 * makes mballoc fail assertions later.
4500 start = max(start, rounddown(ac->ac_o_ex.fe_logical,
4501 (ext4_lblk_t)EXT4_BLOCKS_PER_GROUP(ac->ac_sb)));
4503 /* avoid unnecessary preallocation that may trigger assertions */
4504 if (start + size > EXT_MAX_BLOCKS)
4505 size = EXT_MAX_BLOCKS - start;
4507 /* don't cover already allocated blocks in selected range */
4508 if (ar->pleft && start <= ar->lleft) {
4509 size -= ar->lleft + 1 - start;
4510 start = ar->lleft + 1;
4512 if (ar->pright && start + size - 1 >= ar->lright)
4513 size -= start + size - ar->lright;
4516 * Trim allocation request for filesystems with artificially small
4519 if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb))
4520 size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb);
4524 ext4_mb_pa_adjust_overlap(ac, &start, &end);
4529 * In this function "start" and "size" are normalized for better
4530 * alignment and length such that we could preallocate more blocks.
4531 * This normalization is done such that original request of
4532 * ac->ac_o_ex.fe_logical & fe_len should always lie within "start" and
4533 * "size" boundaries.
4534 * (Note fe_len can be relaxed since FS block allocation API does not
4535 * provide gurantee on number of contiguous blocks allocation since that
4536 * depends upon free space left, etc).
4537 * In case of inode pa, later we use the allocated blocks
4538 * [pa_pstart + fe_logical - pa_lstart, fe_len/size] from the preallocated
4539 * range of goal/best blocks [start, size] to put it at the
4540 * ac_o_ex.fe_logical extent of this inode.
4541 * (See ext4_mb_use_inode_pa() for more details)
4543 if (start + size <= ac->ac_o_ex.fe_logical ||
4544 start > ac->ac_o_ex.fe_logical) {
4545 ext4_msg(ac->ac_sb, KERN_ERR,
4546 "start %lu, size %lu, fe_logical %lu",
4547 (unsigned long) start, (unsigned long) size,
4548 (unsigned long) ac->ac_o_ex.fe_logical);
4551 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
4553 /* now prepare goal request */
4555 /* XXX: is it better to align blocks WRT to logical
4556 * placement or satisfy big request as is */
4557 ac->ac_g_ex.fe_logical = start;
4558 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
4559 ac->ac_orig_goal_len = ac->ac_g_ex.fe_len;
4561 /* define goal start in order to merge */
4562 if (ar->pright && (ar->lright == (start + size)) &&
4563 ar->pright >= size &&
4564 ar->pright - size >= le32_to_cpu(es->s_first_data_block)) {
4565 /* merge to the right */
4566 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
4567 &ac->ac_g_ex.fe_group,
4568 &ac->ac_g_ex.fe_start);
4569 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
4571 if (ar->pleft && (ar->lleft + 1 == start) &&
4572 ar->pleft + 1 < ext4_blocks_count(es)) {
4573 /* merge to the left */
4574 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
4575 &ac->ac_g_ex.fe_group,
4576 &ac->ac_g_ex.fe_start);
4577 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
4580 mb_debug(ac->ac_sb, "goal: %lld(was %lld) blocks at %u\n", size,
4584 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
4586 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4588 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len >= 1) {
4589 atomic_inc(&sbi->s_bal_reqs);
4590 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
4591 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
4592 atomic_inc(&sbi->s_bal_success);
4594 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
4595 for (int i=0; i<EXT4_MB_NUM_CRS; i++) {
4596 atomic_add(ac->ac_cX_found[i], &sbi->s_bal_cX_ex_scanned[i]);
4599 atomic_add(ac->ac_groups_scanned, &sbi->s_bal_groups_scanned);
4600 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
4601 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
4602 atomic_inc(&sbi->s_bal_goals);
4603 /* did we allocate as much as normalizer originally wanted? */
4604 if (ac->ac_f_ex.fe_len == ac->ac_orig_goal_len)
4605 atomic_inc(&sbi->s_bal_len_goals);
4607 if (ac->ac_found > sbi->s_mb_max_to_scan)
4608 atomic_inc(&sbi->s_bal_breaks);
4611 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
4612 trace_ext4_mballoc_alloc(ac);
4614 trace_ext4_mballoc_prealloc(ac);
4618 * Called on failure; free up any blocks from the inode PA for this
4619 * context. We don't need this for MB_GROUP_PA because we only change
4620 * pa_free in ext4_mb_release_context(), but on failure, we've already
4621 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
4623 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
4625 struct ext4_prealloc_space *pa = ac->ac_pa;
4626 struct ext4_buddy e4b;
4630 if (ac->ac_f_ex.fe_len == 0)
4632 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
4633 if (WARN_RATELIMIT(err,
4634 "ext4: mb_load_buddy failed (%d)", err))
4636 * This should never happen since we pin the
4637 * pages in the ext4_allocation_context so
4638 * ext4_mb_load_buddy() should never fail.
4641 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
4642 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
4643 ac->ac_f_ex.fe_len);
4644 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
4645 ext4_mb_unload_buddy(&e4b);
4648 if (pa->pa_type == MB_INODE_PA) {
4649 spin_lock(&pa->pa_lock);
4650 pa->pa_free += ac->ac_b_ex.fe_len;
4651 spin_unlock(&pa->pa_lock);
4656 * use blocks preallocated to inode
4658 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
4659 struct ext4_prealloc_space *pa)
4661 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4666 /* found preallocated blocks, use them */
4667 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
4668 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
4669 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
4670 len = EXT4_NUM_B2C(sbi, end - start);
4671 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
4672 &ac->ac_b_ex.fe_start);
4673 ac->ac_b_ex.fe_len = len;
4674 ac->ac_status = AC_STATUS_FOUND;
4677 BUG_ON(start < pa->pa_pstart);
4678 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
4679 BUG_ON(pa->pa_free < len);
4680 BUG_ON(ac->ac_b_ex.fe_len <= 0);
4683 mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa);
4687 * use blocks preallocated to locality group
4689 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
4690 struct ext4_prealloc_space *pa)
4692 unsigned int len = ac->ac_o_ex.fe_len;
4694 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
4695 &ac->ac_b_ex.fe_group,
4696 &ac->ac_b_ex.fe_start);
4697 ac->ac_b_ex.fe_len = len;
4698 ac->ac_status = AC_STATUS_FOUND;
4701 /* we don't correct pa_pstart or pa_len here to avoid
4702 * possible race when the group is being loaded concurrently
4703 * instead we correct pa later, after blocks are marked
4704 * in on-disk bitmap -- see ext4_mb_release_context()
4705 * Other CPUs are prevented from allocating from this pa by lg_mutex
4707 mb_debug(ac->ac_sb, "use %u/%u from group pa %p\n",
4708 pa->pa_lstart, len, pa);
4712 * Return the prealloc space that have minimal distance
4713 * from the goal block. @cpa is the prealloc
4714 * space that is having currently known minimal distance
4715 * from the goal block.
4717 static struct ext4_prealloc_space *
4718 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
4719 struct ext4_prealloc_space *pa,
4720 struct ext4_prealloc_space *cpa)
4722 ext4_fsblk_t cur_distance, new_distance;
4725 atomic_inc(&pa->pa_count);
4728 cur_distance = abs(goal_block - cpa->pa_pstart);
4729 new_distance = abs(goal_block - pa->pa_pstart);
4731 if (cur_distance <= new_distance)
4734 /* drop the previous reference */
4735 atomic_dec(&cpa->pa_count);
4736 atomic_inc(&pa->pa_count);
4741 * check if found pa meets EXT4_MB_HINT_GOAL_ONLY
4744 ext4_mb_pa_goal_check(struct ext4_allocation_context *ac,
4745 struct ext4_prealloc_space *pa)
4747 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4750 if (likely(!(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)))
4754 * If EXT4_MB_HINT_GOAL_ONLY is set, ac_g_ex will not be adjusted
4755 * in ext4_mb_normalize_request and will keep same with ac_o_ex
4756 * from ext4_mb_initialize_context. Choose ac_g_ex here to keep
4757 * consistent with ext4_mb_find_by_goal.
4759 start = pa->pa_pstart +
4760 (ac->ac_g_ex.fe_logical - pa->pa_lstart);
4761 if (ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex) != start)
4764 if (ac->ac_g_ex.fe_len > pa->pa_len -
4765 EXT4_B2C(sbi, ac->ac_g_ex.fe_logical - pa->pa_lstart))
4772 * search goal blocks in preallocated space
4774 static noinline_for_stack bool
4775 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
4777 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4779 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
4780 struct ext4_locality_group *lg;
4781 struct ext4_prealloc_space *tmp_pa = NULL, *cpa = NULL;
4782 struct rb_node *iter;
4783 ext4_fsblk_t goal_block;
4785 /* only data can be preallocated */
4786 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4790 * first, try per-file preallocation by searching the inode pa rbtree.
4792 * Here, we can't do a direct traversal of the tree because
4793 * ext4_mb_discard_group_preallocation() can paralelly mark the pa
4794 * deleted and that can cause direct traversal to skip some entries.
4796 read_lock(&ei->i_prealloc_lock);
4798 if (RB_EMPTY_ROOT(&ei->i_prealloc_node)) {
4803 * Step 1: Find a pa with logical start immediately adjacent to the
4804 * original logical start. This could be on the left or right.
4806 * (tmp_pa->pa_lstart never changes so we can skip locking for it).
4808 for (iter = ei->i_prealloc_node.rb_node; iter;
4809 iter = ext4_mb_pa_rb_next_iter(ac->ac_o_ex.fe_logical,
4810 tmp_pa->pa_lstart, iter)) {
4811 tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
4812 pa_node.inode_node);
4816 * Step 2: The adjacent pa might be to the right of logical start, find
4817 * the left adjacent pa. After this step we'd have a valid tmp_pa whose
4818 * logical start is towards the left of original request's logical start
4820 if (tmp_pa->pa_lstart > ac->ac_o_ex.fe_logical) {
4821 struct rb_node *tmp;
4822 tmp = rb_prev(&tmp_pa->pa_node.inode_node);
4825 tmp_pa = rb_entry(tmp, struct ext4_prealloc_space,
4826 pa_node.inode_node);
4829 * If there is no adjacent pa to the left then finding
4830 * an overlapping pa is not possible hence stop searching
4837 BUG_ON(!(tmp_pa && tmp_pa->pa_lstart <= ac->ac_o_ex.fe_logical));
4840 * Step 3: If the left adjacent pa is deleted, keep moving left to find
4841 * the first non deleted adjacent pa. After this step we should have a
4842 * valid tmp_pa which is guaranteed to be non deleted.
4844 for (iter = &tmp_pa->pa_node.inode_node;; iter = rb_prev(iter)) {
4847 * no non deleted left adjacent pa, so stop searching
4852 tmp_pa = rb_entry(iter, struct ext4_prealloc_space,
4853 pa_node.inode_node);
4854 spin_lock(&tmp_pa->pa_lock);
4855 if (tmp_pa->pa_deleted == 0) {
4857 * We will keep holding the pa_lock from
4858 * this point on because we don't want group discard
4859 * to delete this pa underneath us. Since group
4860 * discard is anyways an ENOSPC operation it
4861 * should be okay for it to wait a few more cycles.
4865 spin_unlock(&tmp_pa->pa_lock);
4869 BUG_ON(!(tmp_pa && tmp_pa->pa_lstart <= ac->ac_o_ex.fe_logical));
4870 BUG_ON(tmp_pa->pa_deleted == 1);
4873 * Step 4: We now have the non deleted left adjacent pa. Only this
4874 * pa can possibly satisfy the request hence check if it overlaps
4875 * original logical start and stop searching if it doesn't.
4877 if (ac->ac_o_ex.fe_logical >= pa_logical_end(sbi, tmp_pa)) {
4878 spin_unlock(&tmp_pa->pa_lock);
4882 /* non-extent files can't have physical blocks past 2^32 */
4883 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
4884 (tmp_pa->pa_pstart + EXT4_C2B(sbi, tmp_pa->pa_len) >
4885 EXT4_MAX_BLOCK_FILE_PHYS)) {
4887 * Since PAs don't overlap, we won't find any other PA to
4890 spin_unlock(&tmp_pa->pa_lock);
4894 if (tmp_pa->pa_free && likely(ext4_mb_pa_goal_check(ac, tmp_pa))) {
4895 atomic_inc(&tmp_pa->pa_count);
4896 ext4_mb_use_inode_pa(ac, tmp_pa);
4897 spin_unlock(&tmp_pa->pa_lock);
4898 read_unlock(&ei->i_prealloc_lock);
4902 * We found a valid overlapping pa but couldn't use it because
4903 * it had no free blocks. This should ideally never happen
4906 * 1. When a new inode pa is added to rbtree it must have
4907 * pa_free > 0 since otherwise we won't actually need
4910 * 2. An inode pa that is in the rbtree can only have it's
4911 * pa_free become zero when another thread calls:
4912 * ext4_mb_new_blocks
4913 * ext4_mb_use_preallocated
4914 * ext4_mb_use_inode_pa
4916 * 3. Further, after the above calls make pa_free == 0, we will
4917 * immediately remove it from the rbtree in:
4918 * ext4_mb_new_blocks
4919 * ext4_mb_release_context
4922 * 4. Since the pa_free becoming 0 and pa_free getting removed
4923 * from tree both happen in ext4_mb_new_blocks, which is always
4924 * called with i_data_sem held for data allocations, we can be
4925 * sure that another process will never see a pa in rbtree with
4928 WARN_ON_ONCE(tmp_pa->pa_free == 0);
4930 spin_unlock(&tmp_pa->pa_lock);
4932 read_unlock(&ei->i_prealloc_lock);
4934 /* can we use group allocation? */
4935 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
4938 /* inode may have no locality group for some reason */
4942 order = fls(ac->ac_o_ex.fe_len) - 1;
4943 if (order > PREALLOC_TB_SIZE - 1)
4944 /* The max size of hash table is PREALLOC_TB_SIZE */
4945 order = PREALLOC_TB_SIZE - 1;
4947 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
4949 * search for the prealloc space that is having
4950 * minimal distance from the goal block.
4952 for (i = order; i < PREALLOC_TB_SIZE; i++) {
4954 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[i],
4956 spin_lock(&tmp_pa->pa_lock);
4957 if (tmp_pa->pa_deleted == 0 &&
4958 tmp_pa->pa_free >= ac->ac_o_ex.fe_len) {
4960 cpa = ext4_mb_check_group_pa(goal_block,
4963 spin_unlock(&tmp_pa->pa_lock);
4968 ext4_mb_use_group_pa(ac, cpa);
4975 * the function goes through all preallocation in this group and marks them
4976 * used in in-core bitmap. buddy must be generated from this bitmap
4977 * Need to be called with ext4 group lock held
4979 static noinline_for_stack
4980 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
4983 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
4984 struct ext4_prealloc_space *pa;
4985 struct list_head *cur;
4986 ext4_group_t groupnr;
4987 ext4_grpblk_t start;
4988 int preallocated = 0;
4994 /* all form of preallocation discards first load group,
4995 * so the only competing code is preallocation use.
4996 * we don't need any locking here
4997 * notice we do NOT ignore preallocations with pa_deleted
4998 * otherwise we could leave used blocks available for
4999 * allocation in buddy when concurrent ext4_mb_put_pa()
5000 * is dropping preallocation
5002 list_for_each(cur, &grp->bb_prealloc_list) {
5003 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
5004 spin_lock(&pa->pa_lock);
5005 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
5008 spin_unlock(&pa->pa_lock);
5009 if (unlikely(len == 0))
5011 BUG_ON(groupnr != group);
5012 mb_set_bits(bitmap, start, len);
5013 preallocated += len;
5015 mb_debug(sb, "preallocated %d for group %u\n", preallocated, group);
5018 static void ext4_mb_mark_pa_deleted(struct super_block *sb,
5019 struct ext4_prealloc_space *pa)
5021 struct ext4_inode_info *ei;
5023 if (pa->pa_deleted) {
5024 ext4_warning(sb, "deleted pa, type:%d, pblk:%llu, lblk:%u, len:%d\n",
5025 pa->pa_type, pa->pa_pstart, pa->pa_lstart,
5032 if (pa->pa_type == MB_INODE_PA) {
5033 ei = EXT4_I(pa->pa_inode);
5034 atomic_dec(&ei->i_prealloc_active);
5038 static inline void ext4_mb_pa_free(struct ext4_prealloc_space *pa)
5041 BUG_ON(atomic_read(&pa->pa_count));
5042 BUG_ON(pa->pa_deleted == 0);
5043 kmem_cache_free(ext4_pspace_cachep, pa);
5046 static void ext4_mb_pa_callback(struct rcu_head *head)
5048 struct ext4_prealloc_space *pa;
5050 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
5051 ext4_mb_pa_free(pa);
5055 * drops a reference to preallocated space descriptor
5056 * if this was the last reference and the space is consumed
5058 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
5059 struct super_block *sb, struct ext4_prealloc_space *pa)
5062 ext4_fsblk_t grp_blk;
5063 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
5065 /* in this short window concurrent discard can set pa_deleted */
5066 spin_lock(&pa->pa_lock);
5067 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
5068 spin_unlock(&pa->pa_lock);
5072 if (pa->pa_deleted == 1) {
5073 spin_unlock(&pa->pa_lock);
5077 ext4_mb_mark_pa_deleted(sb, pa);
5078 spin_unlock(&pa->pa_lock);
5080 grp_blk = pa->pa_pstart;
5082 * If doing group-based preallocation, pa_pstart may be in the
5083 * next group when pa is used up
5085 if (pa->pa_type == MB_GROUP_PA)
5088 grp = ext4_get_group_number(sb, grp_blk);
5093 * P1 (buddy init) P2 (regular allocation)
5094 * find block B in PA
5095 * copy on-disk bitmap to buddy
5096 * mark B in on-disk bitmap
5097 * drop PA from group
5098 * mark all PAs in buddy
5100 * thus, P1 initializes buddy with B available. to prevent this
5101 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
5104 ext4_lock_group(sb, grp);
5105 list_del(&pa->pa_group_list);
5106 ext4_unlock_group(sb, grp);
5108 if (pa->pa_type == MB_INODE_PA) {
5109 write_lock(pa->pa_node_lock.inode_lock);
5110 rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node);
5111 write_unlock(pa->pa_node_lock.inode_lock);
5112 ext4_mb_pa_free(pa);
5114 spin_lock(pa->pa_node_lock.lg_lock);
5115 list_del_rcu(&pa->pa_node.lg_list);
5116 spin_unlock(pa->pa_node_lock.lg_lock);
5117 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
5121 static void ext4_mb_pa_rb_insert(struct rb_root *root, struct rb_node *new)
5123 struct rb_node **iter = &root->rb_node, *parent = NULL;
5124 struct ext4_prealloc_space *iter_pa, *new_pa;
5125 ext4_lblk_t iter_start, new_start;
5128 iter_pa = rb_entry(*iter, struct ext4_prealloc_space,
5129 pa_node.inode_node);
5130 new_pa = rb_entry(new, struct ext4_prealloc_space,
5131 pa_node.inode_node);
5132 iter_start = iter_pa->pa_lstart;
5133 new_start = new_pa->pa_lstart;
5136 if (new_start < iter_start)
5137 iter = &((*iter)->rb_left);
5139 iter = &((*iter)->rb_right);
5142 rb_link_node(new, parent, iter);
5143 rb_insert_color(new, root);
5147 * creates new preallocated space for given inode
5149 static noinline_for_stack void
5150 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
5152 struct super_block *sb = ac->ac_sb;
5153 struct ext4_sb_info *sbi = EXT4_SB(sb);
5154 struct ext4_prealloc_space *pa;
5155 struct ext4_group_info *grp;
5156 struct ext4_inode_info *ei;
5158 /* preallocate only when found space is larger then requested */
5159 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
5160 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
5161 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
5162 BUG_ON(ac->ac_pa == NULL);
5166 if (ac->ac_b_ex.fe_len < ac->ac_orig_goal_len) {
5167 struct ext4_free_extent ex = {
5168 .fe_logical = ac->ac_g_ex.fe_logical,
5169 .fe_len = ac->ac_orig_goal_len,
5171 loff_t orig_goal_end = extent_logical_end(sbi, &ex);
5173 /* we can't allocate as much as normalizer wants.
5174 * so, found space must get proper lstart
5175 * to cover original request */
5176 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
5177 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
5180 * Use the below logic for adjusting best extent as it keeps
5181 * fragmentation in check while ensuring logical range of best
5182 * extent doesn't overflow out of goal extent:
5184 * 1. Check if best ex can be kept at end of goal (before
5185 * cr_best_avail trimmed it) and still cover original start
5186 * 2. Else, check if best ex can be kept at start of goal and
5187 * still cover original start
5188 * 3. Else, keep the best ex at start of original request.
5190 ex.fe_len = ac->ac_b_ex.fe_len;
5192 ex.fe_logical = orig_goal_end - EXT4_C2B(sbi, ex.fe_len);
5193 if (ac->ac_o_ex.fe_logical >= ex.fe_logical)
5196 ex.fe_logical = ac->ac_g_ex.fe_logical;
5197 if (ac->ac_o_ex.fe_logical < extent_logical_end(sbi, &ex))
5200 ex.fe_logical = ac->ac_o_ex.fe_logical;
5202 ac->ac_b_ex.fe_logical = ex.fe_logical;
5204 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
5205 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
5206 BUG_ON(extent_logical_end(sbi, &ex) > orig_goal_end);
5209 pa->pa_lstart = ac->ac_b_ex.fe_logical;
5210 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
5211 pa->pa_len = ac->ac_b_ex.fe_len;
5212 pa->pa_free = pa->pa_len;
5213 spin_lock_init(&pa->pa_lock);
5214 INIT_LIST_HEAD(&pa->pa_group_list);
5216 pa->pa_type = MB_INODE_PA;
5218 mb_debug(sb, "new inode pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
5219 pa->pa_len, pa->pa_lstart);
5220 trace_ext4_mb_new_inode_pa(ac, pa);
5222 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
5223 ext4_mb_use_inode_pa(ac, pa);
5225 ei = EXT4_I(ac->ac_inode);
5226 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
5230 pa->pa_node_lock.inode_lock = &ei->i_prealloc_lock;
5231 pa->pa_inode = ac->ac_inode;
5233 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
5235 write_lock(pa->pa_node_lock.inode_lock);
5236 ext4_mb_pa_rb_insert(&ei->i_prealloc_node, &pa->pa_node.inode_node);
5237 write_unlock(pa->pa_node_lock.inode_lock);
5238 atomic_inc(&ei->i_prealloc_active);
5242 * creates new preallocated space for locality group inodes belongs to
5244 static noinline_for_stack void
5245 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
5247 struct super_block *sb = ac->ac_sb;
5248 struct ext4_locality_group *lg;
5249 struct ext4_prealloc_space *pa;
5250 struct ext4_group_info *grp;
5252 /* preallocate only when found space is larger then requested */
5253 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
5254 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
5255 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
5256 BUG_ON(ac->ac_pa == NULL);
5260 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
5261 pa->pa_lstart = pa->pa_pstart;
5262 pa->pa_len = ac->ac_b_ex.fe_len;
5263 pa->pa_free = pa->pa_len;
5264 spin_lock_init(&pa->pa_lock);
5265 INIT_LIST_HEAD(&pa->pa_node.lg_list);
5266 INIT_LIST_HEAD(&pa->pa_group_list);
5268 pa->pa_type = MB_GROUP_PA;
5270 mb_debug(sb, "new group pa %p: %llu/%d for %u\n", pa, pa->pa_pstart,
5271 pa->pa_len, pa->pa_lstart);
5272 trace_ext4_mb_new_group_pa(ac, pa);
5274 ext4_mb_use_group_pa(ac, pa);
5275 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
5277 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
5283 pa->pa_node_lock.lg_lock = &lg->lg_prealloc_lock;
5284 pa->pa_inode = NULL;
5286 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
5289 * We will later add the new pa to the right bucket
5290 * after updating the pa_free in ext4_mb_release_context
5294 static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
5296 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
5297 ext4_mb_new_group_pa(ac);
5299 ext4_mb_new_inode_pa(ac);
5303 * finds all unused blocks in on-disk bitmap, frees them in
5304 * in-core bitmap and buddy.
5305 * @pa must be unlinked from inode and group lists, so that
5306 * nobody else can find/use it.
5307 * the caller MUST hold group/inode locks.
5308 * TODO: optimize the case when there are no in-core structures yet
5310 static noinline_for_stack void
5311 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
5312 struct ext4_prealloc_space *pa)
5314 struct super_block *sb = e4b->bd_sb;
5315 struct ext4_sb_info *sbi = EXT4_SB(sb);
5320 unsigned long long grp_blk_start;
5323 BUG_ON(pa->pa_deleted == 0);
5324 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
5325 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
5326 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
5327 end = bit + pa->pa_len;
5330 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
5333 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
5334 mb_debug(sb, "free preallocated %u/%u in group %u\n",
5335 (unsigned) ext4_group_first_block_no(sb, group) + bit,
5336 (unsigned) next - bit, (unsigned) group);
5339 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
5340 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
5341 EXT4_C2B(sbi, bit)),
5343 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
5346 if (free != pa->pa_free) {
5347 ext4_msg(e4b->bd_sb, KERN_CRIT,
5348 "pa %p: logic %lu, phys. %lu, len %d",
5349 pa, (unsigned long) pa->pa_lstart,
5350 (unsigned long) pa->pa_pstart,
5352 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
5355 * pa is already deleted so we use the value obtained
5356 * from the bitmap and continue.
5359 atomic_add(free, &sbi->s_mb_discarded);
5362 static noinline_for_stack void
5363 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
5364 struct ext4_prealloc_space *pa)
5366 struct super_block *sb = e4b->bd_sb;
5370 trace_ext4_mb_release_group_pa(sb, pa);
5371 BUG_ON(pa->pa_deleted == 0);
5372 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
5373 if (unlikely(group != e4b->bd_group && pa->pa_len != 0)) {
5374 ext4_warning(sb, "bad group: expected %u, group %u, pa_start %llu",
5375 e4b->bd_group, group, pa->pa_pstart);
5378 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
5379 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
5380 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
5384 * releases all preallocations in given group
5386 * first, we need to decide discard policy:
5387 * - when do we discard
5389 * - how many do we discard
5390 * 1) how many requested
5392 static noinline_for_stack int
5393 ext4_mb_discard_group_preallocations(struct super_block *sb,
5394 ext4_group_t group, int *busy)
5396 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
5397 struct buffer_head *bitmap_bh = NULL;
5398 struct ext4_prealloc_space *pa, *tmp;
5400 struct ext4_buddy e4b;
5401 struct ext4_inode_info *ei;
5407 mb_debug(sb, "discard preallocation for group %u\n", group);
5408 if (list_empty(&grp->bb_prealloc_list))
5411 bitmap_bh = ext4_read_block_bitmap(sb, group);
5412 if (IS_ERR(bitmap_bh)) {
5413 err = PTR_ERR(bitmap_bh);
5414 ext4_error_err(sb, -err,
5415 "Error %d reading block bitmap for %u",
5420 err = ext4_mb_load_buddy(sb, group, &e4b);
5422 ext4_warning(sb, "Error %d loading buddy information for %u",
5428 ext4_lock_group(sb, group);
5429 list_for_each_entry_safe(pa, tmp,
5430 &grp->bb_prealloc_list, pa_group_list) {
5431 spin_lock(&pa->pa_lock);
5432 if (atomic_read(&pa->pa_count)) {
5433 spin_unlock(&pa->pa_lock);
5437 if (pa->pa_deleted) {
5438 spin_unlock(&pa->pa_lock);
5442 /* seems this one can be freed ... */
5443 ext4_mb_mark_pa_deleted(sb, pa);
5446 this_cpu_inc(discard_pa_seq);
5448 /* we can trust pa_free ... */
5449 free += pa->pa_free;
5451 spin_unlock(&pa->pa_lock);
5453 list_del(&pa->pa_group_list);
5454 list_add(&pa->u.pa_tmp_list, &list);
5457 /* now free all selected PAs */
5458 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
5460 /* remove from object (inode or locality group) */
5461 if (pa->pa_type == MB_GROUP_PA) {
5462 spin_lock(pa->pa_node_lock.lg_lock);
5463 list_del_rcu(&pa->pa_node.lg_list);
5464 spin_unlock(pa->pa_node_lock.lg_lock);
5466 write_lock(pa->pa_node_lock.inode_lock);
5467 ei = EXT4_I(pa->pa_inode);
5468 rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node);
5469 write_unlock(pa->pa_node_lock.inode_lock);
5472 list_del(&pa->u.pa_tmp_list);
5474 if (pa->pa_type == MB_GROUP_PA) {
5475 ext4_mb_release_group_pa(&e4b, pa);
5476 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
5478 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
5479 ext4_mb_pa_free(pa);
5483 ext4_unlock_group(sb, group);
5484 ext4_mb_unload_buddy(&e4b);
5487 mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n",
5488 free, group, grp->bb_free);
5493 * releases all non-used preallocated blocks for given inode
5495 * It's important to discard preallocations under i_data_sem
5496 * We don't want another block to be served from the prealloc
5497 * space when we are discarding the inode prealloc space.
5499 * FIXME!! Make sure it is valid at all the call sites
5501 void ext4_discard_preallocations(struct inode *inode)
5503 struct ext4_inode_info *ei = EXT4_I(inode);
5504 struct super_block *sb = inode->i_sb;
5505 struct buffer_head *bitmap_bh = NULL;
5506 struct ext4_prealloc_space *pa, *tmp;
5507 ext4_group_t group = 0;
5509 struct ext4_buddy e4b;
5510 struct rb_node *iter;
5513 if (!S_ISREG(inode->i_mode))
5516 if (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY)
5519 mb_debug(sb, "discard preallocation for inode %lu\n",
5521 trace_ext4_discard_preallocations(inode,
5522 atomic_read(&ei->i_prealloc_active));
5525 /* first, collect all pa's in the inode */
5526 write_lock(&ei->i_prealloc_lock);
5527 for (iter = rb_first(&ei->i_prealloc_node); iter;
5528 iter = rb_next(iter)) {
5529 pa = rb_entry(iter, struct ext4_prealloc_space,
5530 pa_node.inode_node);
5531 BUG_ON(pa->pa_node_lock.inode_lock != &ei->i_prealloc_lock);
5533 spin_lock(&pa->pa_lock);
5534 if (atomic_read(&pa->pa_count)) {
5535 /* this shouldn't happen often - nobody should
5536 * use preallocation while we're discarding it */
5537 spin_unlock(&pa->pa_lock);
5538 write_unlock(&ei->i_prealloc_lock);
5539 ext4_msg(sb, KERN_ERR,
5540 "uh-oh! used pa while discarding");
5542 schedule_timeout_uninterruptible(HZ);
5546 if (pa->pa_deleted == 0) {
5547 ext4_mb_mark_pa_deleted(sb, pa);
5548 spin_unlock(&pa->pa_lock);
5549 rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node);
5550 list_add(&pa->u.pa_tmp_list, &list);
5554 /* someone is deleting pa right now */
5555 spin_unlock(&pa->pa_lock);
5556 write_unlock(&ei->i_prealloc_lock);
5558 /* we have to wait here because pa_deleted
5559 * doesn't mean pa is already unlinked from
5560 * the list. as we might be called from
5561 * ->clear_inode() the inode will get freed
5562 * and concurrent thread which is unlinking
5563 * pa from inode's list may access already
5564 * freed memory, bad-bad-bad */
5566 /* XXX: if this happens too often, we can
5567 * add a flag to force wait only in case
5568 * of ->clear_inode(), but not in case of
5569 * regular truncate */
5570 schedule_timeout_uninterruptible(HZ);
5573 write_unlock(&ei->i_prealloc_lock);
5575 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
5576 BUG_ON(pa->pa_type != MB_INODE_PA);
5577 group = ext4_get_group_number(sb, pa->pa_pstart);
5579 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
5580 GFP_NOFS|__GFP_NOFAIL);
5582 ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
5587 bitmap_bh = ext4_read_block_bitmap(sb, group);
5588 if (IS_ERR(bitmap_bh)) {
5589 err = PTR_ERR(bitmap_bh);
5590 ext4_error_err(sb, -err, "Error %d reading block bitmap for %u",
5592 ext4_mb_unload_buddy(&e4b);
5596 ext4_lock_group(sb, group);
5597 list_del(&pa->pa_group_list);
5598 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
5599 ext4_unlock_group(sb, group);
5601 ext4_mb_unload_buddy(&e4b);
5604 list_del(&pa->u.pa_tmp_list);
5605 ext4_mb_pa_free(pa);
5609 static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac)
5611 struct ext4_prealloc_space *pa;
5613 BUG_ON(ext4_pspace_cachep == NULL);
5614 pa = kmem_cache_zalloc(ext4_pspace_cachep, GFP_NOFS);
5617 atomic_set(&pa->pa_count, 1);
5622 static void ext4_mb_pa_put_free(struct ext4_allocation_context *ac)
5624 struct ext4_prealloc_space *pa = ac->ac_pa;
5628 WARN_ON(!atomic_dec_and_test(&pa->pa_count));
5630 * current function is only called due to an error or due to
5631 * len of found blocks < len of requested blocks hence the PA has not
5632 * been added to grp->bb_prealloc_list. So we don't need to lock it
5635 ext4_mb_pa_free(pa);
5638 #ifdef CONFIG_EXT4_DEBUG
5639 static inline void ext4_mb_show_pa(struct super_block *sb)
5641 ext4_group_t i, ngroups;
5643 if (ext4_forced_shutdown(sb))
5646 ngroups = ext4_get_groups_count(sb);
5647 mb_debug(sb, "groups: ");
5648 for (i = 0; i < ngroups; i++) {
5649 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
5650 struct ext4_prealloc_space *pa;
5651 ext4_grpblk_t start;
5652 struct list_head *cur;
5656 ext4_lock_group(sb, i);
5657 list_for_each(cur, &grp->bb_prealloc_list) {
5658 pa = list_entry(cur, struct ext4_prealloc_space,
5660 spin_lock(&pa->pa_lock);
5661 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
5663 spin_unlock(&pa->pa_lock);
5664 mb_debug(sb, "PA:%u:%d:%d\n", i, start,
5667 ext4_unlock_group(sb, i);
5668 mb_debug(sb, "%u: %d/%d\n", i, grp->bb_free,
5673 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
5675 struct super_block *sb = ac->ac_sb;
5677 if (ext4_forced_shutdown(sb))
5680 mb_debug(sb, "Can't allocate:"
5681 " Allocation context details:");
5682 mb_debug(sb, "status %u flags 0x%x",
5683 ac->ac_status, ac->ac_flags);
5684 mb_debug(sb, "orig %lu/%lu/%lu@%lu, "
5685 "goal %lu/%lu/%lu@%lu, "
5686 "best %lu/%lu/%lu@%lu cr %d",
5687 (unsigned long)ac->ac_o_ex.fe_group,
5688 (unsigned long)ac->ac_o_ex.fe_start,
5689 (unsigned long)ac->ac_o_ex.fe_len,
5690 (unsigned long)ac->ac_o_ex.fe_logical,
5691 (unsigned long)ac->ac_g_ex.fe_group,
5692 (unsigned long)ac->ac_g_ex.fe_start,
5693 (unsigned long)ac->ac_g_ex.fe_len,
5694 (unsigned long)ac->ac_g_ex.fe_logical,
5695 (unsigned long)ac->ac_b_ex.fe_group,
5696 (unsigned long)ac->ac_b_ex.fe_start,
5697 (unsigned long)ac->ac_b_ex.fe_len,
5698 (unsigned long)ac->ac_b_ex.fe_logical,
5699 (int)ac->ac_criteria);
5700 mb_debug(sb, "%u found", ac->ac_found);
5701 mb_debug(sb, "used pa: %s, ", ac->ac_pa ? "yes" : "no");
5703 mb_debug(sb, "pa_type %s\n", ac->ac_pa->pa_type == MB_GROUP_PA ?
5704 "group pa" : "inode pa");
5705 ext4_mb_show_pa(sb);
5708 static inline void ext4_mb_show_pa(struct super_block *sb)
5711 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
5713 ext4_mb_show_pa(ac->ac_sb);
5718 * We use locality group preallocation for small size file. The size of the
5719 * file is determined by the current size or the resulting size after
5720 * allocation which ever is larger
5722 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
5724 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
5726 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
5727 int bsbits = ac->ac_sb->s_blocksize_bits;
5729 bool inode_pa_eligible, group_pa_eligible;
5731 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
5734 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
5737 group_pa_eligible = sbi->s_mb_group_prealloc > 0;
5738 inode_pa_eligible = true;
5739 size = extent_logical_end(sbi, &ac->ac_o_ex);
5740 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
5743 /* No point in using inode preallocation for closed files */
5744 if ((size == isize) && !ext4_fs_is_busy(sbi) &&
5745 !inode_is_open_for_write(ac->ac_inode))
5746 inode_pa_eligible = false;
5748 size = max(size, isize);
5749 /* Don't use group allocation for large files */
5750 if (size > sbi->s_mb_stream_request)
5751 group_pa_eligible = false;
5753 if (!group_pa_eligible) {
5754 if (inode_pa_eligible)
5755 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
5757 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
5761 BUG_ON(ac->ac_lg != NULL);
5763 * locality group prealloc space are per cpu. The reason for having
5764 * per cpu locality group is to reduce the contention between block
5765 * request from multiple CPUs.
5767 ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
5769 /* we're going to use group allocation */
5770 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
5772 /* serialize all allocations in the group */
5773 mutex_lock(&ac->ac_lg->lg_mutex);
5776 static noinline_for_stack void
5777 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
5778 struct ext4_allocation_request *ar)
5780 struct super_block *sb = ar->inode->i_sb;
5781 struct ext4_sb_info *sbi = EXT4_SB(sb);
5782 struct ext4_super_block *es = sbi->s_es;
5786 ext4_grpblk_t block;
5788 /* we can't allocate > group size */
5791 /* just a dirty hack to filter too big requests */
5792 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
5793 len = EXT4_CLUSTERS_PER_GROUP(sb);
5795 /* start searching from the goal */
5797 if (goal < le32_to_cpu(es->s_first_data_block) ||
5798 goal >= ext4_blocks_count(es))
5799 goal = le32_to_cpu(es->s_first_data_block);
5800 ext4_get_group_no_and_offset(sb, goal, &group, &block);
5802 /* set up allocation goals */
5803 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
5804 ac->ac_status = AC_STATUS_CONTINUE;
5806 ac->ac_inode = ar->inode;
5807 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
5808 ac->ac_o_ex.fe_group = group;
5809 ac->ac_o_ex.fe_start = block;
5810 ac->ac_o_ex.fe_len = len;
5811 ac->ac_g_ex = ac->ac_o_ex;
5812 ac->ac_orig_goal_len = ac->ac_g_ex.fe_len;
5813 ac->ac_flags = ar->flags;
5815 /* we have to define context: we'll work with a file or
5816 * locality group. this is a policy, actually */
5817 ext4_mb_group_or_file(ac);
5819 mb_debug(sb, "init ac: %u blocks @ %u, goal %u, flags 0x%x, 2^%d, "
5820 "left: %u/%u, right %u/%u to %swritable\n",
5821 (unsigned) ar->len, (unsigned) ar->logical,
5822 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
5823 (unsigned) ar->lleft, (unsigned) ar->pleft,
5824 (unsigned) ar->lright, (unsigned) ar->pright,
5825 inode_is_open_for_write(ar->inode) ? "" : "non-");
5828 static noinline_for_stack void
5829 ext4_mb_discard_lg_preallocations(struct super_block *sb,
5830 struct ext4_locality_group *lg,
5831 int order, int total_entries)
5833 ext4_group_t group = 0;
5834 struct ext4_buddy e4b;
5835 LIST_HEAD(discard_list);
5836 struct ext4_prealloc_space *pa, *tmp;
5838 mb_debug(sb, "discard locality group preallocation\n");
5840 spin_lock(&lg->lg_prealloc_lock);
5841 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
5843 lockdep_is_held(&lg->lg_prealloc_lock)) {
5844 spin_lock(&pa->pa_lock);
5845 if (atomic_read(&pa->pa_count)) {
5847 * This is the pa that we just used
5848 * for block allocation. So don't
5851 spin_unlock(&pa->pa_lock);
5854 if (pa->pa_deleted) {
5855 spin_unlock(&pa->pa_lock);
5858 /* only lg prealloc space */
5859 BUG_ON(pa->pa_type != MB_GROUP_PA);
5861 /* seems this one can be freed ... */
5862 ext4_mb_mark_pa_deleted(sb, pa);
5863 spin_unlock(&pa->pa_lock);
5865 list_del_rcu(&pa->pa_node.lg_list);
5866 list_add(&pa->u.pa_tmp_list, &discard_list);
5869 if (total_entries <= 5) {
5871 * we want to keep only 5 entries
5872 * allowing it to grow to 8. This
5873 * mak sure we don't call discard
5874 * soon for this list.
5879 spin_unlock(&lg->lg_prealloc_lock);
5881 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
5884 group = ext4_get_group_number(sb, pa->pa_pstart);
5885 err = ext4_mb_load_buddy_gfp(sb, group, &e4b,
5886 GFP_NOFS|__GFP_NOFAIL);
5888 ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
5892 ext4_lock_group(sb, group);
5893 list_del(&pa->pa_group_list);
5894 ext4_mb_release_group_pa(&e4b, pa);
5895 ext4_unlock_group(sb, group);
5897 ext4_mb_unload_buddy(&e4b);
5898 list_del(&pa->u.pa_tmp_list);
5899 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
5904 * We have incremented pa_count. So it cannot be freed at this
5905 * point. Also we hold lg_mutex. So no parallel allocation is
5906 * possible from this lg. That means pa_free cannot be updated.
5908 * A parallel ext4_mb_discard_group_preallocations is possible.
5909 * which can cause the lg_prealloc_list to be updated.
5912 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
5914 int order, added = 0, lg_prealloc_count = 1;
5915 struct super_block *sb = ac->ac_sb;
5916 struct ext4_locality_group *lg = ac->ac_lg;
5917 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
5919 order = fls(pa->pa_free) - 1;
5920 if (order > PREALLOC_TB_SIZE - 1)
5921 /* The max size of hash table is PREALLOC_TB_SIZE */
5922 order = PREALLOC_TB_SIZE - 1;
5923 /* Add the prealloc space to lg */
5924 spin_lock(&lg->lg_prealloc_lock);
5925 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
5927 lockdep_is_held(&lg->lg_prealloc_lock)) {
5928 spin_lock(&tmp_pa->pa_lock);
5929 if (tmp_pa->pa_deleted) {
5930 spin_unlock(&tmp_pa->pa_lock);
5933 if (!added && pa->pa_free < tmp_pa->pa_free) {
5934 /* Add to the tail of the previous entry */
5935 list_add_tail_rcu(&pa->pa_node.lg_list,
5936 &tmp_pa->pa_node.lg_list);
5939 * we want to count the total
5940 * number of entries in the list
5943 spin_unlock(&tmp_pa->pa_lock);
5944 lg_prealloc_count++;
5947 list_add_tail_rcu(&pa->pa_node.lg_list,
5948 &lg->lg_prealloc_list[order]);
5949 spin_unlock(&lg->lg_prealloc_lock);
5951 /* Now trim the list to be not more than 8 elements */
5952 if (lg_prealloc_count > 8)
5953 ext4_mb_discard_lg_preallocations(sb, lg,
5954 order, lg_prealloc_count);
5958 * release all resource we used in allocation
5960 static void ext4_mb_release_context(struct ext4_allocation_context *ac)
5962 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
5963 struct ext4_prealloc_space *pa = ac->ac_pa;
5965 if (pa->pa_type == MB_GROUP_PA) {
5966 /* see comment in ext4_mb_use_group_pa() */
5967 spin_lock(&pa->pa_lock);
5968 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
5969 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
5970 pa->pa_free -= ac->ac_b_ex.fe_len;
5971 pa->pa_len -= ac->ac_b_ex.fe_len;
5972 spin_unlock(&pa->pa_lock);
5975 * We want to add the pa to the right bucket.
5976 * Remove it from the list and while adding
5977 * make sure the list to which we are adding
5980 if (likely(pa->pa_free)) {
5981 spin_lock(pa->pa_node_lock.lg_lock);
5982 list_del_rcu(&pa->pa_node.lg_list);
5983 spin_unlock(pa->pa_node_lock.lg_lock);
5984 ext4_mb_add_n_trim(ac);
5988 ext4_mb_put_pa(ac, ac->ac_sb, pa);
5990 if (ac->ac_bitmap_page)
5991 put_page(ac->ac_bitmap_page);
5992 if (ac->ac_buddy_page)
5993 put_page(ac->ac_buddy_page);
5994 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
5995 mutex_unlock(&ac->ac_lg->lg_mutex);
5996 ext4_mb_collect_stats(ac);
5999 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
6001 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
6003 int freed = 0, busy = 0;
6006 trace_ext4_mb_discard_preallocations(sb, needed);
6009 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
6011 for (i = 0; i < ngroups && needed > 0; i++) {
6012 ret = ext4_mb_discard_group_preallocations(sb, i, &busy);
6018 if (needed > 0 && busy && ++retry < 3) {
6026 static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb,
6027 struct ext4_allocation_context *ac, u64 *seq)
6033 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
6038 seq_retry = ext4_get_discard_pa_seq_sum();
6039 if (!(ac->ac_flags & EXT4_MB_STRICT_CHECK) || seq_retry != *seq) {
6040 ac->ac_flags |= EXT4_MB_STRICT_CHECK;
6046 mb_debug(sb, "freed %d, retry ? %s\n", freed, ret ? "yes" : "no");
6051 * Simple allocator for Ext4 fast commit replay path. It searches for blocks
6052 * linearly starting at the goal block and also excludes the blocks which
6053 * are going to be in use after fast commit replay.
6056 ext4_mb_new_blocks_simple(struct ext4_allocation_request *ar, int *errp)
6058 struct buffer_head *bitmap_bh;
6059 struct super_block *sb = ar->inode->i_sb;
6060 struct ext4_sb_info *sbi = EXT4_SB(sb);
6061 ext4_group_t group, nr;
6062 ext4_grpblk_t blkoff;
6063 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
6064 ext4_grpblk_t i = 0;
6065 ext4_fsblk_t goal, block;
6066 struct ext4_super_block *es = sbi->s_es;
6069 if (goal < le32_to_cpu(es->s_first_data_block) ||
6070 goal >= ext4_blocks_count(es))
6071 goal = le32_to_cpu(es->s_first_data_block);
6074 ext4_get_group_no_and_offset(sb, goal, &group, &blkoff);
6075 for (nr = ext4_get_groups_count(sb); nr > 0; nr--) {
6076 bitmap_bh = ext4_read_block_bitmap(sb, group);
6077 if (IS_ERR(bitmap_bh)) {
6078 *errp = PTR_ERR(bitmap_bh);
6079 pr_warn("Failed to read block bitmap\n");
6084 i = mb_find_next_zero_bit(bitmap_bh->b_data, max,
6088 if (ext4_fc_replay_check_excluded(sb,
6089 ext4_group_first_block_no(sb, group) +
6090 EXT4_C2B(sbi, i))) {
6099 if (++group >= ext4_get_groups_count(sb))
6110 block = ext4_group_first_block_no(sb, group) + EXT4_C2B(sbi, i);
6111 ext4_mb_mark_bb(sb, block, 1, true);
6118 * Main entry point into mballoc to allocate blocks
6119 * it tries to use preallocation first, then falls back
6120 * to usual allocation
6122 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
6123 struct ext4_allocation_request *ar, int *errp)
6125 struct ext4_allocation_context *ac = NULL;
6126 struct ext4_sb_info *sbi;
6127 struct super_block *sb;
6128 ext4_fsblk_t block = 0;
6129 unsigned int inquota = 0;
6130 unsigned int reserv_clstrs = 0;
6135 sb = ar->inode->i_sb;
6138 trace_ext4_request_blocks(ar);
6139 if (sbi->s_mount_state & EXT4_FC_REPLAY)
6140 return ext4_mb_new_blocks_simple(ar, errp);
6142 /* Allow to use superuser reservation for quota file */
6143 if (ext4_is_quota_file(ar->inode))
6144 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
6146 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
6147 /* Without delayed allocation we need to verify
6148 * there is enough free blocks to do block allocation
6149 * and verify allocation doesn't exceed the quota limits.
6152 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
6154 /* let others to free the space */
6156 ar->len = ar->len >> 1;
6159 ext4_mb_show_pa(sb);
6163 reserv_clstrs = ar->len;
6164 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
6165 dquot_alloc_block_nofail(ar->inode,
6166 EXT4_C2B(sbi, ar->len));
6169 dquot_alloc_block(ar->inode,
6170 EXT4_C2B(sbi, ar->len))) {
6172 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
6183 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
6190 ext4_mb_initialize_context(ac, ar);
6192 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
6193 seq = this_cpu_read(discard_pa_seq);
6194 if (!ext4_mb_use_preallocated(ac)) {
6195 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
6196 ext4_mb_normalize_request(ac, ar);
6198 *errp = ext4_mb_pa_alloc(ac);
6202 /* allocate space in core */
6203 *errp = ext4_mb_regular_allocator(ac);
6205 * pa allocated above is added to grp->bb_prealloc_list only
6206 * when we were able to allocate some block i.e. when
6207 * ac->ac_status == AC_STATUS_FOUND.
6208 * And error from above mean ac->ac_status != AC_STATUS_FOUND
6209 * So we have to free this pa here itself.
6212 ext4_mb_pa_put_free(ac);
6213 ext4_discard_allocated_blocks(ac);
6216 if (ac->ac_status == AC_STATUS_FOUND &&
6217 ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len)
6218 ext4_mb_pa_put_free(ac);
6220 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
6221 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
6223 ext4_discard_allocated_blocks(ac);
6226 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
6227 ar->len = ac->ac_b_ex.fe_len;
6230 if (++retries < 3 &&
6231 ext4_mb_discard_preallocations_should_retry(sb, ac, &seq))
6234 * If block allocation fails then the pa allocated above
6235 * needs to be freed here itself.
6237 ext4_mb_pa_put_free(ac);
6243 ac->ac_b_ex.fe_len = 0;
6245 ext4_mb_show_ac(ac);
6247 ext4_mb_release_context(ac);
6248 kmem_cache_free(ext4_ac_cachep, ac);
6250 if (inquota && ar->len < inquota)
6251 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
6253 if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
6254 /* release all the reserved blocks if non delalloc */
6255 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
6259 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
6265 * We can merge two free data extents only if the physical blocks
6266 * are contiguous, AND the extents were freed by the same transaction,
6267 * AND the blocks are associated with the same group.
6269 static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi,
6270 struct ext4_free_data *entry,
6271 struct ext4_free_data *new_entry,
6272 struct rb_root *entry_rb_root)
6274 if ((entry->efd_tid != new_entry->efd_tid) ||
6275 (entry->efd_group != new_entry->efd_group))
6277 if (entry->efd_start_cluster + entry->efd_count ==
6278 new_entry->efd_start_cluster) {
6279 new_entry->efd_start_cluster = entry->efd_start_cluster;
6280 new_entry->efd_count += entry->efd_count;
6281 } else if (new_entry->efd_start_cluster + new_entry->efd_count ==
6282 entry->efd_start_cluster) {
6283 new_entry->efd_count += entry->efd_count;
6286 spin_lock(&sbi->s_md_lock);
6287 list_del(&entry->efd_list);
6288 spin_unlock(&sbi->s_md_lock);
6289 rb_erase(&entry->efd_node, entry_rb_root);
6290 kmem_cache_free(ext4_free_data_cachep, entry);
6293 static noinline_for_stack void
6294 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
6295 struct ext4_free_data *new_entry)
6297 ext4_group_t group = e4b->bd_group;
6298 ext4_grpblk_t cluster;
6299 ext4_grpblk_t clusters = new_entry->efd_count;
6300 struct ext4_free_data *entry;
6301 struct ext4_group_info *db = e4b->bd_info;
6302 struct super_block *sb = e4b->bd_sb;
6303 struct ext4_sb_info *sbi = EXT4_SB(sb);
6304 struct rb_node **n = &db->bb_free_root.rb_node, *node;
6305 struct rb_node *parent = NULL, *new_node;
6307 BUG_ON(!ext4_handle_valid(handle));
6308 BUG_ON(e4b->bd_bitmap_page == NULL);
6309 BUG_ON(e4b->bd_buddy_page == NULL);
6311 new_node = &new_entry->efd_node;
6312 cluster = new_entry->efd_start_cluster;
6315 /* first free block exent. We need to
6316 protect buddy cache from being freed,
6317 * otherwise we'll refresh it from
6318 * on-disk bitmap and lose not-yet-available
6320 get_page(e4b->bd_buddy_page);
6321 get_page(e4b->bd_bitmap_page);
6325 entry = rb_entry(parent, struct ext4_free_data, efd_node);
6326 if (cluster < entry->efd_start_cluster)
6328 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
6329 n = &(*n)->rb_right;
6331 ext4_grp_locked_error(sb, group, 0,
6332 ext4_group_first_block_no(sb, group) +
6333 EXT4_C2B(sbi, cluster),
6334 "Block already on to-be-freed list");
6335 kmem_cache_free(ext4_free_data_cachep, new_entry);
6340 rb_link_node(new_node, parent, n);
6341 rb_insert_color(new_node, &db->bb_free_root);
6343 /* Now try to see the extent can be merged to left and right */
6344 node = rb_prev(new_node);
6346 entry = rb_entry(node, struct ext4_free_data, efd_node);
6347 ext4_try_merge_freed_extent(sbi, entry, new_entry,
6348 &(db->bb_free_root));
6351 node = rb_next(new_node);
6353 entry = rb_entry(node, struct ext4_free_data, efd_node);
6354 ext4_try_merge_freed_extent(sbi, entry, new_entry,
6355 &(db->bb_free_root));
6358 spin_lock(&sbi->s_md_lock);
6359 list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list[new_entry->efd_tid & 1]);
6360 sbi->s_mb_free_pending += clusters;
6361 spin_unlock(&sbi->s_md_lock);
6364 static void ext4_free_blocks_simple(struct inode *inode, ext4_fsblk_t block,
6365 unsigned long count)
6367 struct super_block *sb = inode->i_sb;
6369 ext4_grpblk_t blkoff;
6371 ext4_get_group_no_and_offset(sb, block, &group, &blkoff);
6372 ext4_mb_mark_context(NULL, sb, false, group, blkoff, count,
6373 EXT4_MB_BITMAP_MARKED_CHECK |
6374 EXT4_MB_SYNC_UPDATE,
6379 * ext4_mb_clear_bb() -- helper function for freeing blocks.
6380 * Used by ext4_free_blocks()
6381 * @handle: handle for this transaction
6383 * @block: starting physical block to be freed
6384 * @count: number of blocks to be freed
6385 * @flags: flags used by ext4_free_blocks
6387 static void ext4_mb_clear_bb(handle_t *handle, struct inode *inode,
6388 ext4_fsblk_t block, unsigned long count,
6391 struct super_block *sb = inode->i_sb;
6392 struct ext4_group_info *grp;
6393 unsigned int overflow;
6395 ext4_group_t block_group;
6396 struct ext4_sb_info *sbi;
6397 struct ext4_buddy e4b;
6398 unsigned int count_clusters;
6401 ext4_grpblk_t changed;
6405 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
6406 !ext4_inode_block_valid(inode, block, count)) {
6407 ext4_error(sb, "Freeing blocks in system zone - "
6408 "Block = %llu, count = %lu", block, count);
6409 /* err = 0. ext4_std_error should be a no op */
6412 flags |= EXT4_FREE_BLOCKS_VALIDATED;
6416 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
6418 grp = ext4_get_group_info(sb, block_group);
6419 if (unlikely(!grp || EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
6423 * Check to see if we are freeing blocks across a group
6426 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
6427 overflow = EXT4_C2B(sbi, bit) + count -
6428 EXT4_BLOCKS_PER_GROUP(sb);
6430 /* The range changed so it's no longer validated */
6431 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6433 count_clusters = EXT4_NUM_B2C(sbi, count);
6434 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
6436 /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
6437 err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b,
6438 GFP_NOFS|__GFP_NOFAIL);
6442 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
6443 !ext4_inode_block_valid(inode, block, count)) {
6444 ext4_error(sb, "Freeing blocks in system zone - "
6445 "Block = %llu, count = %lu", block, count);
6446 /* err = 0. ext4_std_error should be a no op */
6450 #ifdef AGGRESSIVE_CHECK
6451 mark_flags |= EXT4_MB_BITMAP_MARKED_CHECK;
6453 err = ext4_mb_mark_context(handle, sb, false, block_group, bit,
6454 count_clusters, mark_flags, &changed);
6457 if (err && changed == 0)
6460 #ifdef AGGRESSIVE_CHECK
6461 BUG_ON(changed != count_clusters);
6465 * We need to make sure we don't reuse the freed block until after the
6466 * transaction is committed. We make an exception if the inode is to be
6467 * written in writeback mode since writeback mode has weak data
6468 * consistency guarantees.
6470 if (ext4_handle_valid(handle) &&
6471 ((flags & EXT4_FREE_BLOCKS_METADATA) ||
6472 !ext4_should_writeback_data(inode))) {
6473 struct ext4_free_data *new_entry;
6475 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
6478 new_entry = kmem_cache_alloc(ext4_free_data_cachep,
6479 GFP_NOFS|__GFP_NOFAIL);
6480 new_entry->efd_start_cluster = bit;
6481 new_entry->efd_group = block_group;
6482 new_entry->efd_count = count_clusters;
6483 new_entry->efd_tid = handle->h_transaction->t_tid;
6485 ext4_lock_group(sb, block_group);
6486 ext4_mb_free_metadata(handle, &e4b, new_entry);
6488 if (test_opt(sb, DISCARD)) {
6489 err = ext4_issue_discard(sb, block_group, bit,
6490 count_clusters, NULL);
6491 if (err && err != -EOPNOTSUPP)
6492 ext4_msg(sb, KERN_WARNING, "discard request in"
6493 " group:%u block:%d count:%lu failed"
6494 " with %d", block_group, bit, count,
6497 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
6499 ext4_lock_group(sb, block_group);
6500 mb_free_blocks(inode, &e4b, bit, count_clusters);
6503 ext4_unlock_group(sb, block_group);
6506 * on a bigalloc file system, defer the s_freeclusters_counter
6507 * update to the caller (ext4_remove_space and friends) so they
6508 * can determine if a cluster freed here should be rereserved
6510 if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) {
6511 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
6512 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
6513 percpu_counter_add(&sbi->s_freeclusters_counter,
6517 if (overflow && !err) {
6520 ext4_mb_unload_buddy(&e4b);
6521 /* The range changed so it's no longer validated */
6522 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6527 ext4_mb_unload_buddy(&e4b);
6529 ext4_std_error(sb, err);
6533 * ext4_free_blocks() -- Free given blocks and update quota
6534 * @handle: handle for this transaction
6536 * @bh: optional buffer of the block to be freed
6537 * @block: starting physical block to be freed
6538 * @count: number of blocks to be freed
6539 * @flags: flags used by ext4_free_blocks
6541 void ext4_free_blocks(handle_t *handle, struct inode *inode,
6542 struct buffer_head *bh, ext4_fsblk_t block,
6543 unsigned long count, int flags)
6545 struct super_block *sb = inode->i_sb;
6546 unsigned int overflow;
6547 struct ext4_sb_info *sbi;
6553 BUG_ON(block != bh->b_blocknr);
6555 block = bh->b_blocknr;
6558 if (sbi->s_mount_state & EXT4_FC_REPLAY) {
6559 ext4_free_blocks_simple(inode, block, EXT4_NUM_B2C(sbi, count));
6565 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
6566 !ext4_inode_block_valid(inode, block, count)) {
6567 ext4_error(sb, "Freeing blocks not in datazone - "
6568 "block = %llu, count = %lu", block, count);
6571 flags |= EXT4_FREE_BLOCKS_VALIDATED;
6573 ext4_debug("freeing block %llu\n", block);
6574 trace_ext4_free_blocks(inode, block, count, flags);
6576 if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
6579 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
6584 * If the extent to be freed does not begin on a cluster
6585 * boundary, we need to deal with partial clusters at the
6586 * beginning and end of the extent. Normally we will free
6587 * blocks at the beginning or the end unless we are explicitly
6588 * requested to avoid doing so.
6590 overflow = EXT4_PBLK_COFF(sbi, block);
6592 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
6593 overflow = sbi->s_cluster_ratio - overflow;
6595 if (count > overflow)
6603 /* The range changed so it's no longer validated */
6604 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6606 overflow = EXT4_LBLK_COFF(sbi, count);
6608 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
6609 if (count > overflow)
6614 count += sbi->s_cluster_ratio - overflow;
6615 /* The range changed so it's no longer validated */
6616 flags &= ~EXT4_FREE_BLOCKS_VALIDATED;
6619 if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
6621 int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA;
6623 for (i = 0; i < count; i++) {
6626 bh = sb_find_get_block(inode->i_sb, block + i);
6627 ext4_forget(handle, is_metadata, inode, bh, block + i);
6631 ext4_mb_clear_bb(handle, inode, block, count, flags);
6635 * ext4_group_add_blocks() -- Add given blocks to an existing group
6636 * @handle: handle to this transaction
6638 * @block: start physical block to add to the block group
6639 * @count: number of blocks to free
6641 * This marks the blocks as free in the bitmap and buddy.
6643 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
6644 ext4_fsblk_t block, unsigned long count)
6646 ext4_group_t block_group;
6648 struct ext4_sb_info *sbi = EXT4_SB(sb);
6649 struct ext4_buddy e4b;
6651 ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block);
6652 ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1);
6653 unsigned long cluster_count = last_cluster - first_cluster + 1;
6654 ext4_grpblk_t changed;
6656 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
6658 if (cluster_count == 0)
6661 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
6663 * Check to see if we are freeing blocks across a group
6666 if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) {
6667 ext4_warning(sb, "too many blocks added to group %u",
6673 err = ext4_mb_load_buddy(sb, block_group, &e4b);
6677 if (!ext4_sb_block_valid(sb, NULL, block, count)) {
6678 ext4_error(sb, "Adding blocks in system zones - "
6679 "Block = %llu, count = %lu",
6685 err = ext4_mb_mark_context(handle, sb, false, block_group, bit,
6686 cluster_count, EXT4_MB_BITMAP_MARKED_CHECK,
6688 if (err && changed == 0)
6691 if (changed != cluster_count)
6692 ext4_error(sb, "bit already cleared in group %u", block_group);
6694 ext4_lock_group(sb, block_group);
6695 mb_free_blocks(NULL, &e4b, bit, cluster_count);
6696 ext4_unlock_group(sb, block_group);
6697 percpu_counter_add(&sbi->s_freeclusters_counter,
6701 ext4_mb_unload_buddy(&e4b);
6703 ext4_std_error(sb, err);
6708 * ext4_trim_extent -- function to TRIM one single free extent in the group
6709 * @sb: super block for the file system
6710 * @start: starting block of the free extent in the alloc. group
6711 * @count: number of blocks to TRIM
6712 * @e4b: ext4 buddy for the group
6714 * Trim "count" blocks starting at "start" in the "group". To assure that no
6715 * one will allocate those blocks, mark it as used in buddy bitmap. This must
6716 * be called with under the group lock.
6718 static int ext4_trim_extent(struct super_block *sb,
6719 int start, int count, struct ext4_buddy *e4b)
6723 struct ext4_free_extent ex;
6724 ext4_group_t group = e4b->bd_group;
6727 trace_ext4_trim_extent(sb, group, start, count);
6729 assert_spin_locked(ext4_group_lock_ptr(sb, group));
6731 ex.fe_start = start;
6732 ex.fe_group = group;
6736 * Mark blocks used, so no one can reuse them while
6739 mb_mark_used(e4b, &ex);
6740 ext4_unlock_group(sb, group);
6741 ret = ext4_issue_discard(sb, group, start, count, NULL);
6742 ext4_lock_group(sb, group);
6743 mb_free_blocks(NULL, e4b, start, ex.fe_len);
6747 static ext4_grpblk_t ext4_last_grp_cluster(struct super_block *sb,
6750 unsigned long nr_clusters_in_group;
6752 if (grp < (ext4_get_groups_count(sb) - 1))
6753 nr_clusters_in_group = EXT4_CLUSTERS_PER_GROUP(sb);
6755 nr_clusters_in_group = (ext4_blocks_count(EXT4_SB(sb)->s_es) -
6756 ext4_group_first_block_no(sb, grp))
6757 >> EXT4_CLUSTER_BITS(sb);
6759 return nr_clusters_in_group - 1;
6762 static bool ext4_trim_interrupted(void)
6764 return fatal_signal_pending(current) || freezing(current);
6767 static int ext4_try_to_trim_range(struct super_block *sb,
6768 struct ext4_buddy *e4b, ext4_grpblk_t start,
6769 ext4_grpblk_t max, ext4_grpblk_t minblocks)
6770 __acquires(ext4_group_lock_ptr(sb, e4b->bd_group))
6771 __releases(ext4_group_lock_ptr(sb, e4b->bd_group))
6773 ext4_grpblk_t next, count, free_count, last, origin_start;
6774 bool set_trimmed = false;
6777 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
6780 last = ext4_last_grp_cluster(sb, e4b->bd_group);
6781 bitmap = e4b->bd_bitmap;
6782 if (start == 0 && max >= last)
6784 origin_start = start;
6785 start = max(e4b->bd_info->bb_first_free, start);
6789 while (start <= max) {
6790 start = mb_find_next_zero_bit(bitmap, max + 1, start);
6794 next = mb_find_next_bit(bitmap, last + 1, start);
6795 if (origin_start == 0 && next >= last)
6798 if ((next - start) >= minblocks) {
6799 int ret = ext4_trim_extent(sb, start, next - start, e4b);
6801 if (ret && ret != -EOPNOTSUPP)
6803 count += next - start;
6805 free_count += next - start;
6808 if (ext4_trim_interrupted())
6811 if (need_resched()) {
6812 ext4_unlock_group(sb, e4b->bd_group);
6814 ext4_lock_group(sb, e4b->bd_group);
6817 if ((e4b->bd_info->bb_free - free_count) < minblocks)
6822 EXT4_MB_GRP_SET_TRIMMED(e4b->bd_info);
6828 * ext4_trim_all_free -- function to trim all free space in alloc. group
6829 * @sb: super block for file system
6830 * @group: group to be trimmed
6831 * @start: first group block to examine
6832 * @max: last group block to examine
6833 * @minblocks: minimum extent block count
6835 * ext4_trim_all_free walks through group's block bitmap searching for free
6836 * extents. When the free extent is found, mark it as used in group buddy
6837 * bitmap. Then issue a TRIM command on this extent and free the extent in
6838 * the group buddy bitmap.
6840 static ext4_grpblk_t
6841 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
6842 ext4_grpblk_t start, ext4_grpblk_t max,
6843 ext4_grpblk_t minblocks)
6845 struct ext4_buddy e4b;
6848 trace_ext4_trim_all_free(sb, group, start, max);
6850 ret = ext4_mb_load_buddy(sb, group, &e4b);
6852 ext4_warning(sb, "Error %d loading buddy information for %u",
6857 ext4_lock_group(sb, group);
6859 if (!EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) ||
6860 minblocks < EXT4_SB(sb)->s_last_trim_minblks)
6861 ret = ext4_try_to_trim_range(sb, &e4b, start, max, minblocks);
6865 ext4_unlock_group(sb, group);
6866 ext4_mb_unload_buddy(&e4b);
6868 ext4_debug("trimmed %d blocks in the group %d\n",
6875 * ext4_trim_fs() -- trim ioctl handle function
6876 * @sb: superblock for filesystem
6877 * @range: fstrim_range structure
6879 * start: First Byte to trim
6880 * len: number of Bytes to trim from start
6881 * minlen: minimum extent length in Bytes
6882 * ext4_trim_fs goes through all allocation groups containing Bytes from
6883 * start to start+len. For each such a group ext4_trim_all_free function
6884 * is invoked to trim all free space.
6886 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
6888 unsigned int discard_granularity = bdev_discard_granularity(sb->s_bdev);
6889 struct ext4_group_info *grp;
6890 ext4_group_t group, first_group, last_group;
6891 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
6892 uint64_t start, end, minlen, trimmed = 0;
6893 ext4_fsblk_t first_data_blk =
6894 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
6895 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
6898 start = range->start >> sb->s_blocksize_bits;
6899 end = start + (range->len >> sb->s_blocksize_bits) - 1;
6900 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
6901 range->minlen >> sb->s_blocksize_bits);
6903 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
6904 start >= max_blks ||
6905 range->len < sb->s_blocksize)
6907 /* No point to try to trim less than discard granularity */
6908 if (range->minlen < discard_granularity) {
6909 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
6910 discard_granularity >> sb->s_blocksize_bits);
6911 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb))
6914 if (end >= max_blks - 1)
6916 if (end <= first_data_blk)
6918 if (start < first_data_blk)
6919 start = first_data_blk;
6921 /* Determine first and last group to examine based on start and end */
6922 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
6923 &first_group, &first_cluster);
6924 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
6925 &last_group, &last_cluster);
6927 /* end now represents the last cluster to discard in this group */
6928 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
6930 for (group = first_group; group <= last_group; group++) {
6931 if (ext4_trim_interrupted())
6933 grp = ext4_get_group_info(sb, group);
6936 /* We only do this if the grp has never been initialized */
6937 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
6938 ret = ext4_mb_init_group(sb, group, GFP_NOFS);
6944 * For all the groups except the last one, last cluster will
6945 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
6946 * change it for the last group, note that last_cluster is
6947 * already computed earlier by ext4_get_group_no_and_offset()
6949 if (group == last_group)
6951 if (grp->bb_free >= minlen) {
6952 cnt = ext4_trim_all_free(sb, group, first_cluster,
6962 * For every group except the first one, we are sure
6963 * that the first cluster to discard will be cluster #0.
6969 EXT4_SB(sb)->s_last_trim_minblks = minlen;
6972 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
6976 /* Iterate all the free extents in the group. */
6978 ext4_mballoc_query_range(
6979 struct super_block *sb,
6981 ext4_grpblk_t start,
6983 ext4_mballoc_query_range_fn formatter,
6988 struct ext4_buddy e4b;
6991 error = ext4_mb_load_buddy(sb, group, &e4b);
6994 bitmap = e4b.bd_bitmap;
6996 ext4_lock_group(sb, group);
6998 start = max(e4b.bd_info->bb_first_free, start);
6999 if (end >= EXT4_CLUSTERS_PER_GROUP(sb))
7000 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
7002 while (start <= end) {
7003 start = mb_find_next_zero_bit(bitmap, end + 1, start);
7006 next = mb_find_next_bit(bitmap, end + 1, start);
7008 ext4_unlock_group(sb, group);
7009 error = formatter(sb, group, start, next - start, priv);
7012 ext4_lock_group(sb, group);
7017 ext4_unlock_group(sb, group);
7019 ext4_mb_unload_buddy(&e4b);
7024 #ifdef CONFIG_EXT4_KUNIT_TESTS
7025 #include "mballoc-test.c"